WO2023082787A1

WO2023082787A1 - Method and apparatus for determining contribution degree of participant in federated learning, and federated learning training method and apparatus

Info

Publication number: WO2023082787A1
Application number: PCT/CN2022/116570
Authority: WO
Inventors: 杨程屹; 刘嘉; 李增祥
Original assignee: 新智我来网络科技有限公司
Priority date: 2021-11-10
Filing date: 2022-09-01
Publication date: 2023-05-19

Abstract

A method and apparatus for determining a contribution degree of a participant in federated learning, and a federated learning training method and apparatus. The method for determining the contribution degree of the participant in federated learning comprises: constructing all participant combinations, and calculating the weight of each participant combination; determining a utility change value of a federated learning model before and after an aggregation period, establishing a lookup table, and according to the utility change value, determining whether to calculate a contribution value of each participant; selecting one participant combination from all participant combinations, estimating a marginal contribution value of the participant, and according to an estimation result and the weight, determining whether to calculate the utility value of the participant combination by using an interpolation function; and updating the lookup table according to the utility value of the participant combination, sequentially selecting the participant combinations until the utility values of all participant combinations are calculated, using the utility values of the participant combinations to update the lookup table, and on the basis of a finally updated lookup table, calculating the contribution value of the participant.

Description

Method for determining participant contribution in joint learning, joint learning training method and device

technical field

The present disclosure relates to the technical field of joint learning, and in particular to a method for determining the contribution degree of a participant in joint learning, a joint learning training method and a device.

Background technique

With the development of artificial intelligence and distributed machine learning technology, the joint learning method of machine learning by combining different participants has become a mainstream trend in training artificial intelligence models. As a new type of distributed machine learning framework, federated learning meets the needs of multiple clients for model training under the requirements of data security.

In the prior art, when measuring the contribution of the participants in the federated learning, the federated learning system is usually used to determine the contribution of each participant to the federated learning model in the federated learning based on the amount of local data of each participant. However, because the local data of each participant may have problems such as uneven data quality and different forms of local data, the existing calculation method of joint learning contribution has low calculation accuracy, inaccurate calculation results, and large amount of calculation data. Relatively large, low computational efficiency and other issues.

In view of the above-mentioned problems in the prior art, it is necessary to provide a method for determining the contribution of joint learning participants that can improve the calculation accuracy of the contribution value, reduce the amount of calculation data, make the calculation result of the contribution value more accurate, and have higher calculation efficiency.

Contents of the invention

In view of this, the embodiments of the present disclosure provide a method for determining the contribution degree of a participant in joint learning, a joint learning training method, and a device to solve the problem of low calculation accuracy in the calculation method of joint learning contribution in the prior art. The calculation results are inaccurate, the amount of calculation data is relatively large, and the calculation efficiency is low.

The first aspect of the embodiments of the present disclosure provides a method for determining the contribution of a participant in joint learning, including:

Determine the participants of the joint learning, construct all participant combinations based on the participants, and calculate the weight corresponding to each participant combination;

Determine the utility change value corresponding to the joint learning model before and after the aggregation period and establish a lookup table, and judge whether to calculate the contribution value of the participants in the aggregation period according to the utility change value;

When the judgment result is yes, select a participant combination from all the participant combinations according to the predetermined order, and estimate the marginal contribution value of each participant in the selected participant combination, according to the estimation result and the participant combination Weight, to determine whether to use the interpolation function to calculate the utility value of the participant combination;

When the judgment result is yes, use the interpolation function to calculate the utility value of the participant combination; when the judgment result is no, use the preset model deduction method to calculate the utility value of the participant combination, and use the calculated The value updates the lookup table;

Select each participant combination in turn until the utility value of all participant combinations is calculated, and use the utility values of all participant combinations to update the lookup table to obtain the final updated lookup table, so that based on the final updated lookup table, Calculate the contribution value of each participant, and determine the contribution of the participant in the joint learning according to the contribution value.

The second aspect of the embodiments of the present disclosure provides an apparatus for determining the contribution of a participant in joint learning, including:

The construction module is configured to determine the participants of the joint learning, construct all participant combinations based on the participants, and calculate the weight corresponding to each participant combination;

The first judging module is configured to determine the utility change value corresponding to the joint learning model before and after the aggregation period and establish a lookup table, and judge whether to calculate the contribution value of the participant in the aggregation period according to the utility change value;

The second judgment module is configured to select a participant combination from all participant combinations in a predetermined order when the judgment result is yes, and estimate the marginal contribution value of each participant in the selected participant combination, According to the estimation result and the weight of the participant combination, judge whether to use the interpolation function to calculate the utility value of the participant combination;

The update module is configured to use an interpolation function to calculate the utility value of the participant combination when the judgment result is yes, and to calculate the utility value of the participant combination by using a preset model derivation method when the judgment result is no, and to calculate the utility value of the participant combination according to the calculated Update the lookup table with the utility value of the participant combination;

The calculation module is configured to select each combination of participants in turn until the utility values of all the combinations of participants are calculated, and use the utility values of all the combinations of participants to update the lookup table to obtain a final updated lookup table, so that based on the final The updated lookup table calculates the contribution value of each participant, and determines the contribution of the participant in the joint learning according to the contribution value.

The third aspect of the embodiments of the present disclosure provides a method for determining the contribution degree of a participant in joint learning, including:

determining all participant combinations based on the participants in the joint learning, and calculating the weight of each participant combination in the all participant combinations;

determining a first utility value of the joint model before the start of the current aggregation period, and a second utility value of the joint model after the end of the current aggregation period, calculating a utility change value based on the first utility value and the second utility value, and Establish a lookup table; wherein, the utility change value is used to determine whether to calculate the contribution value of each participant in the current aggregation period;

When judging and calculating the contribution value, select a participant combination from all the participant combinations, and calculate the marginal contribution value corresponding to each participant in the participant combination, according to the marginal contribution value and the Weight, judging whether to use the first estimation method or the second estimation method to estimate the utility value of the participant combination;

determining the estimation result of the utility value of the participant combination by the first estimation method or the second estimation method, using the estimation result to update the lookup table, and iteratively estimating in turn to obtain each participant combination The utility value of , and obtain the final lookup table obtained after updating according to the utility value, so as to use the final lookup table to calculate the contribution value of the participant.

The fourth aspect of the embodiments of the present disclosure provides an apparatus for determining the contribution of a participant in joint learning, including:

A determination module configured to determine all participant combinations based on the participants in the joint learning, and calculate the weight of each participant combination in the all participant combinations;

A building module configured to determine a first utility value of the joint model before the start of the current aggregation period, and a second utility value of the joint model after the end of the current aggregation period, based on the first utility value and the second utility value Calculating the utility change value, and establishing a lookup table; wherein, the utility change value is used to judge whether to calculate the contribution value of each participant in the current aggregation period;

The judging module is configured to, when judging and calculating the contribution value, select a participant combination from all the participant combinations, and calculate the marginal contribution value corresponding to each participant in the participant combination, according to the The marginal contribution value and the weight determine to use the first estimation method or the second estimation method to estimate the utility value of the participant combination;

The calculation module is configured to determine the estimation result of the utility value of the participant combination by the first estimation method or the second estimation method, use the estimation result to update the look-up table, and iteratively estimate in turn to obtain each A utility value of the participant combination, and a final lookup table obtained after updating according to the utility value is obtained, so as to use the final lookup table to calculate the contribution value of the participant.

The fifth aspect of the embodiments of the present disclosure provides a method for determining the contribution of participants in joint learning, including:

Based on the framework of joint learning, multiple participant groups are generated, and a set of participant groups composed of multiple participant groups is determined, and weights of the participant groups are calculated, wherein each of the participant groups includes at least two participants;

Determine the aggregation period in the joint learning, obtain the utility change value corresponding to the joint learning model before and after the aggregation period and establish a lookup table, and judge whether to calculate the contribution value of each participant in the aggregation period according to the utility change value;

When the judgment result is yes, use the participant groups in the participant group set to randomly generate a full permutation combination, and generate multiple subcombinations according to the order of the participants in the participant group in the full permutation combination, calculate The estimated value of the marginal contribution value when the participant joins the sub-combination, according to the estimated value of the marginal contribution value and the weight of the participant group, it is judged whether to use an interpolation function to add the participant to the sub-combination Calculate the utility value of the new participant group formed;

When the judgment result is yes, use the interpolation function to calculate the utility value of the new participant group, when the judgment result is no, use the preset model derivation method to calculate the utility value of the new participant group, and calculate updating the lookup table with the utility value of the new party group;

Based on the updated lookup table, calculate the marginal contribution value of the participant, and judge whether the marginal contribution value of the participant is converged, and when the judgment result is yes, use the converged marginal contribution value as the participant's Contribution value, when the judgment result is no, generate a new full permutation combination until the calculation of the contribution value of all the participants after convergence, and determine the contribution of the participant in the joint learning according to the contribution value contribution.

The sixth aspect of the embodiments of the present disclosure provides an apparatus for determining the contribution of participants in joint learning, including:

The generation module is configured to generate a plurality of participant groups based on a joint learning architecture, and determine a participant group set composed of a plurality of the participant groups, and calculate the weights of the participant groups, wherein each of the participant groups The party group contains at least two parties;

The establishment module is configured to determine the aggregation period in the joint learning, obtain the utility change value corresponding to the joint learning model before and after the aggregation period, and establish a lookup table, and judge whether to calculate each in the aggregation period according to the utility change value. Contribution value of participants;

The judging module is configured to randomly generate a full permutation combination by using the participant groups in the participant group set when the judgment result is yes, and according to the order of the participants in the participant group in the full permutation combination Generate a plurality of sub-combinations, calculate the estimated value of the marginal contribution value when the participant joins the sub-combination, and judge whether to use an interpolation function to evaluate the participants according to the estimated value of the marginal contribution value and the weight of the participant group The utility value of the new participant group formed after adding the sub-combination is calculated;

The update module is configured to use an interpolation function to calculate the utility value of the new participant group when the judgment result is yes, and to calculate the utility value of the new participant group by using a preset model derivation method when the judgment result is no value, and update the lookup table according to the calculated utility value of the new participant group;

The calculation module is configured to calculate the marginal contribution value of the participant based on the updated lookup table, and judge whether the marginal contribution value of the participant is converged, and when the judgment result is yes, the converged marginal contribution value As the contribution value of the participant, when the judgment result is no, a new full permutation combination is generated until the contribution value of all converged participants is calculated, and the participant is determined according to the contribution value The degree of contribution in the joint learning.

The seventh aspect of the embodiments of the present disclosure provides a joint learning training method, including:

In the current round of aggregation cycle of the joint learning, the local model obtained by the initial model training of the participants of the joint learning according to the local data is obtained, and the aggregation operation is performed on the local models of the participants to obtain the joint model;

Using a preset joint learning contribution value algorithm, calculate the contribution value of each of the participants to the joint model in the current round of aggregation period, and obtain the joint learning contribution value corresponding to each of the participants;

Obtain the initial index of each of the participants, perform a fusion operation on the joint learning contribution value and the initial index, and obtain the contribution index of each of the participants, wherein the initial index is used to represent the contribution of the participant to The initial contribution of joint learning;

calculating the training rounds of the participant in the next aggregation period according to the contribution index, so that the participant can train the local model based on the training rounds in the next aggregation period, Until the training of the joint model reaches the preset target.

The eighth aspect of the embodiments of the present disclosure provides a joint learning and training device, including:

The aggregation module is configured to obtain a local model obtained by the participants of the federated learning through initial model training based on the local data during the current round of the federated learning, and perform an aggregation operation on the local models of the participants to obtain the joint model;

The calculation module is configured to use a preset joint learning contribution value algorithm to calculate the contribution value of each of the participants to the joint model in the current round of aggregation period, and obtain the corresponding joint learning contribution value;

The fusion module is configured to obtain an initial index of each of the participants, perform a fusion operation on the joint learning contribution value and the initial index, and obtain a contribution index of each of the participants, wherein the initial index is used for Characterize the initial contribution of the participants to the joint learning;

A training module configured to calculate the training rounds of the participant in the next round of aggregation period according to the contribution index, so that the participant will, in the next round of aggregation period, based on the training rounds The local model is trained until the training of the joint model reaches a preset target.

The ninth aspect of the embodiments of the present disclosure provides an electronic device, including a memory, a processor, and a computer program stored in the memory and operable on the processor, and the aforementioned method is implemented when the processor executes the program .

A tenth aspect of the embodiments of the present disclosure provides a computer-readable storage medium, where the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the aforementioned method is implemented.

The at least one technical solution adopted in the embodiments of the present disclosure can achieve the following beneficial effects:

By determining the participants of the joint learning, construct all participant combinations based on the participants, and calculate the weight corresponding to each participant combination; determine the utility change value corresponding to the joint learning model before and after the aggregation cycle and establish a lookup table, and judge according to the utility change value Whether to calculate the contribution value of the participants in the aggregation period; when the judgment result is yes, select a participant combination from all the participant combinations according to the predetermined order, and calculate the margin of each participant in the selected participant combination According to the estimation result and the weight of the participant combination, judge whether to use the interpolation function to calculate the utility value of the participant combination; when the judgment result is yes, use the interpolation function to calculate the utility value of the participant combination. When the result is no, use the preset model deduction method to calculate the utility value of the participant combination, and update the lookup table according to the calculated utility value of the participant combination; select each participant combination in turn until all participants are calculated. The utility value of the party combination is used to update the lookup table by using the utility value of all the participants combined to obtain the final updated lookup table, so that the contribution value of each participant can be calculated based on the final updated lookup table, and determined according to the contribution value Contribution of participants in joint learning. The present disclosure can improve the calculation accuracy of the contribution value in joint learning, reduce the amount of calculation data, and make the calculation result of the contribution value more accurate and the calculation efficiency higher.

Description of drawings

In order to more clearly illustrate the technical solutions in the embodiments of the present disclosure, the following will briefly introduce the drawings that need to be used in the embodiments or the description of the prior art. Obviously, the drawings in the following description are only of the present disclosure For some embodiments, those skilled in the art can also obtain other drawings based on these drawings without creative efforts.

FIG. 1 is a schematic diagram of a joint learning architecture provided by an embodiment of the present disclosure;

FIG. 2 is a schematic flowchart of a method for determining the contribution of a participant in joint learning provided by an embodiment of the present disclosure;

Fig. 3 is a schematic flow diagram of a program for calculating a participant's contribution value provided by an embodiment of the present disclosure;

FIG. 4 is a schematic diagram of an apparatus for determining a participant's contribution in joint learning provided by an embodiment of the present disclosure;

FIG. 5 is a schematic flowchart of another method for determining the contribution of a participant in joint learning provided by an embodiment of the present disclosure;

Fig. 6 is a schematic flow diagram of a program for calculating a participant's contribution value provided by an embodiment of the present disclosure;

FIG. 7 is a schematic diagram of another device for determining participant contribution in joint learning provided by an embodiment of the present disclosure;

FIG. 8 is a schematic flowchart of another method for determining the contribution of a participant in joint learning provided by an embodiment of the present disclosure;

FIG. 9 is a schematic flow diagram of a program for calculating participant contribution values provided by an embodiment of the present disclosure;

FIG. 10 is a schematic diagram of another device for determining participant contribution in joint learning provided by an embodiment of the present disclosure;

FIG. 11 is a schematic flowchart of a joint learning training method provided by an embodiment of the present disclosure;

Fig. 12 is a schematic flow chart of calculating the joint learning contribution value of the participants provided by the embodiment of the present disclosure;

Fig. 13 is a schematic diagram of a joint learning and training device provided by an embodiment of the present disclosure;

Fig. 14 is a schematic structural diagram of an electronic device provided by an embodiment of the present disclosure.

Detailed ways

In the following description, for the purpose of illustration rather than limitation, specific details such as specific system structures and techniques are presented for a thorough understanding of the embodiments of the present disclosure. It will be apparent, however, to one skilled in the art that the present disclosure may be practiced in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, devices, circuits, and methods are omitted so as not to obscure the description of the present disclosure with unnecessary detail.

Federated learning refers to the comprehensive utilization of various AI (Artificial Intelligence, artificial intelligence) technologies on the premise of ensuring data security and user privacy, and joint multi-party cooperation to jointly mine data value and generate new intelligent business models and models based on joint modeling. Federated learning has at least the following characteristics:

(1) Participating nodes control the weakly centralized joint training mode of their own data to ensure data privacy and security in the process of co-creating intelligence.

(2) In different application scenarios, use screening and/or combining AI algorithms and privacy-preserving calculations to establish multiple model aggregation optimization strategies to obtain high-level, high-quality models.

(3) On the premise of ensuring data security and user privacy, based on a variety of model aggregation optimization strategies, obtain a performance method to improve the joint learning engine, where the performance method can be solved by solving problems including parallel computing architecture and large-scale cross-domain network Information interaction, intelligent perception, exception handling mechanism, etc., improve the overall performance of the joint learning engine.

(4) Obtain the needs of multi-party users in each scenario, determine and reasonably evaluate the true contribution of each joint participant through the mutual trust mechanism, and distribute incentives.

Based on the above methods, it is possible to establish an AI technology ecology based on joint learning, give full play to the value of industry data, and promote the implementation of scenarios in vertical fields.

At present, with the increase in the number of participants in federated learning and the amount of computing data, how to accurately and quickly evaluate the contribution of each participant in federated learning to the training of the federated learning model has become an urgent problem to be solved. In the existing technology, the joint learning system determines the contribution of each participant to the joint learning model according to the amount of local data of each participant, but because the local data of each participant may have uneven data quality, the format or form of the local data Different, the data features of each participant overlap more and other problems, resulting in low calculation efficiency of the contribution of the joint learning model in the joint learning, and the accuracy of the joint learning contribution of each participant is also low. This will lead to insufficient evaluation of the fairness and impartiality of each participant's contribution when using the joint learning contribution to distribute the benefits of each participant in the later stage.

In view of the above problems in the prior art, it is necessary to provide a method for calculating the contribution of each participant in joint learning based on the Shapley value calculation rule, combined with the marginal contribution value generated when each participant joins the participant combination . Based on the embodiments of the present disclosure, the accurate calculation of the contribution value of each participant's joint learning model training in joint learning can be improved, and the amount of calculation can be reduced, so that the calculation result of the contribution value is more accurate and the calculation efficiency is higher.

FIG. 1 is a schematic diagram of a joint learning architecture provided by an embodiment of the present disclosure. As shown in FIG. 1 , the architecture of joint learning may include a server (central node) 101 , and

participants

102 , 103 , and 104 .

In the joint learning process, the basic model can be established by the server 101, and the server 101 sends the model to the participant 102, the participant 103 and the participant 104 with which a communication connection is established. The basic model can also be uploaded to the server 101 after being created by any participant, and the server 101 sends the model to other participants that have established communication connections with it. Participant 102, participant 103 and participant 104 build a model according to the downloaded basic structure and model parameters, use local data for model training, obtain updated model parameters, and encrypt and upload the updated model parameters to the server 101. The server 101 aggregates the model parameters sent by the participant 102 , the participant 103 and the participant 104 to obtain the global model parameters, and returns the global model parameters to the participant 102 , the participant 103 and the participant 104 . The participant 102, the participant 103 and the participant 104 iterate their models according to the received global model parameters until the models finally converge, thereby realizing the training of the models. In the joint learning process, the data uploaded by participant 102, participant 103, and participant 104 are model parameters, local data will not be uploaded to server 101, and all participants can share the final model parameters, so data can be guaranteed Co-modeling is achieved on the basis of privacy. It should be noted that the number of participants is not limited to the above three, but can be set according to needs, which is not limited in this embodiment of the present disclosure.

Fig. 2 is a schematic flow chart of a method for determining a participant's contribution in joint learning provided by an embodiment of the present disclosure. The method for determining the contribution degree of a participant in the joint learning in FIG. 2 may be executed by a server of the joint learning. As shown in Figure 2, the method for determining the contribution of the participants in the joint learning may specifically include:

S201, determine the participants of the joint learning, construct all participant combinations based on the participants, and calculate the weight corresponding to each participant combination;

S202. Determine the utility change value corresponding to the joint learning model before and after the aggregation period and establish a lookup table, and judge whether to calculate the contribution value of the participant in the aggregation period according to the utility change value;

S203, when the judgment result is yes, select a participant combination from all participant combinations according to a predetermined order, and estimate the marginal contribution value of each participant in the selected participant combination, according to the estimation result and the participant Combination weight, judging whether to use the interpolation function to calculate the utility value of the participant combination;

S204, when the judgment result is yes, use the interpolation function to calculate the utility value of the participant combination; when the judgment result is no, use the preset model deduction method to calculate the utility value of the participant combination, and calculate Update the lookup table with the utility value of ;

S205. Select each participant combination in turn until the utility values of all participant combinations are calculated, and use the utility values of all participant combinations to update the lookup table to obtain the final updated lookup table, so that based on the final updated lookup table table, calculate the contribution value of each participant, and determine the contribution of the participant in the joint learning according to the contribution value.

Specifically, each participant corresponds to a node in the joint learning framework, and each node corresponds to a participant device. The participant device can be a PC, tablet computer, smart phone, smart wearable device, etc., and each participant device There are client terminals of joint learning participants, but the participant devices are not limited to the above-mentioned devices or clients. The federated learning framework also has a node that provides services for the client (that is, the server). The server can be a server for performing aggregation operations. The server can coordinate multiple clients to perform joint learning to obtain a joint learning model. The server may be an independent physical server, or a server cluster or cloud computing server composed of multiple physical servers.

Furthermore, the combination of participants refers to the mutual arrangement and combination of individuals of all participants in the joint learning, and is a combination of participants composed of individuals of the participants. The parties are A, B, and C respectively, then the following combinations of parties can be formed among them: A, B, C, AB, BC, and AC.

Furthermore, in federated learning, an aggregation cycle refers to a round of training for the federated learning model. Each participant client uses local data to train the local model. When the local model training reaches convergence, the trained local model is obtained parameter and send it to the server. All participants will upload their own local model parameters in each aggregation round, and the server will perform a weighted average to obtain a joint model, so each participant will make its own contribution in each round. The round here means The server performs a complete training of the joint learning model.

According to the technical solution provided by the embodiments of the present disclosure, by determining the participants of the joint learning, construct all participant combinations based on the participants, and calculate the weight corresponding to each participant combination; determine the utility change value corresponding to the joint learning model before and after the aggregation period And establish a lookup table, judge whether to calculate the contribution value of the participants in the aggregation period according to the utility change value; when the judgment result is yes, select a participant combination from all the participant combinations in the predetermined order, and calculate According to the estimation result and the weight of the participant combination, judge whether to use the interpolation function to calculate the utility value of the participant combination; when the judgment result is yes, use the interpolation function Calculate the utility value of the participant combination. When the judgment result is no, use the preset model deduction method to calculate the utility value of the participant combination, and update the lookup table according to the calculated utility value of the participant combination; select each A participant combination, until the utility value of all the participant combinations is calculated, the lookup table is updated by using the utility value of all the participant combinations, and the final updated lookup table is obtained, so that based on the final updated lookup table, each The contribution value of the participant, according to the contribution value to determine the contribution of the participant in the joint learning. The present disclosure can improve the calculation accuracy of the contribution value in joint learning, reduce the amount of calculation data, and make the calculation result of the contribution value more accurate and the calculation efficiency higher.

The cyclic process of calculating the contribution value of each participant in the joint learning of the present disclosure will be described in detail below in combination with a specific program flow diagram. FIG. As shown in Figure 3, the program for calculating the contribution value of the participants may specifically include the following:

In some embodiments, constructing all participant combinations based on the participants, and calculating the weight corresponding to each participant combination, includes: constructing all participants according to the number of participants from the fewest to the largest according to all participants in the joint learning Party combination, all participant combinations include multiple participant combinations, and the weight is calculated according to the number of participants in each participant combination; where, the weight is used to represent the probability that the participant combination appears in all the participant combinations.

Specifically, the following will describe in detail the participants of the joint learning and the process of constructing a combination of all participants in combination with a specific embodiment, which may specifically include the following:

Assuming that there are

N participants

1, 2, ... i...n-1, n joint learning, the training has been aggregated for T cycles, and each aggregation cycle t in the training process is recorded, and the local data uploaded by each participant i Model M _i ^(t) , and the joint model M ^(t) after central aggregation, initialize the model M ⁽⁰⁾ , have an evaluation function or utility function V( ) for model performance (such as accuracy, loss, etc.), and the joint Learning model aggregation method Agg(·), thresholds λ, η; where λ represents the first truncated threshold, and η represents the second truncated threshold.

Further, firstly, according to all participants in the joint learning, enumerate all possible participant combinations Ps=[(1,),(2,),(3,)...,(1 ,2),(1,3),(2,3),…P,…N]; for each subcombination S with 0, 1, 2,…n-1 participants, calculate the weight w _{|S |} = |S|! (|N| _- |S|-1)! /|N|! .

It should be noted that each participant combination corresponds to a sub-combination S mentioned above. When calculating the weight of the sub-combination S, it is based on the number of participants in each sub-combination. In a participant combination, a participant Corresponding to the elements in a set, that is, the weight corresponding to the participant combination is calculated according to the number of elements in the participant combination. The weight corresponding to each sub-combination can be considered as the probability of the sub-combination appearing in the overall participant combination.

In some embodiments, determining the utility change value corresponding to the joint learning model before and after the aggregation period and establishing a lookup table includes: for each aggregation period, determining the initial utility value and the final utility value of the aggregation period, and combining the final utility value with the initial utility The difference between the values is used as the utility change value, and a lookup table containing all participant combinations corresponding to the aggregation period is established; the lookup table is initialized so that the lookup table excludes the combination of the empty set participant and the full set participant combination The initial utility value of other participant combinations other than 0; wherein, the lookup table is used to store the utility values corresponding to all participant combinations.

Specifically, calculate the utility value corresponding to the joint learning model at the beginning and after the end of each aggregation period, and establish a lookup table, that is, for each aggregation period, you can first calculate the final utility value of this aggregation period and this aggregation period The calculation process of the initial utility value and the final utility value of the aggregation cycle will be described in detail below in conjunction with a specific embodiment, which may specifically include the following:

For each aggregation period t, calculate v _N =V(M ^(t) ), v ₀ =V(M ^(t-1) ), and establish a lookup table v_lut={():v0,(1,): 0,(2,):0,(3,):0…,(1,2):0,(1,3):0,(2,3):0,…N:vN}, where, v _N indicates the final utility value of the joint model after the end of the current aggregation period, and v ₀ indicates the utility value of the joint model after the previous aggregation period corresponding to the current aggregation period. Of course, v ₀ can also be understood as the current aggregation period before the start of the current aggregation period. For the initial utility value of the period, the difference in different expressions does not constitute a limit to the essential meaning of v ₀ , and the above two expressions are equivalent.

Further, when performing the initialization operation on the lookup table, except for the participant combinations corresponding to the empty set ( ) and the full set N, let the utility values of other participant combinations in all participant combinations Ps be 0. By establishing a v_lut lookup table and using the lookup table to cache the utility value of the participant combination, so as to record the calculated utility value, it is possible to reduce the amount of calculation for subsequent contribution value calculations and avoid repeated calculations.

In some embodiments, judging whether to calculate the contribution value of the participant in the aggregation period according to the utility change value includes: comparing the utility change value of the aggregation period with a preset first cut-off threshold, when the utility change of the aggregation period When the value is less than the first truncation threshold, and the utility change value of two consecutive rounds of aggregation periods is less than the first truncation threshold, it is judged that the contribution value of each participant in the aggregation period is 0; otherwise, the contribution of each participant in the aggregation period value is calculated.

Specifically, by calculating the final utility value of this aggregation period and the initial utility value of this aggregation period, if the difference between the final utility value and the initial utility value corresponding to this aggregation period is less than the first cut-off threshold, and if two consecutive rounds In the aggregation period, when the difference between the final utility value and the initial utility value is less than the first truncation threshold, the calculation ends, and the contribution value of each participant in the current aggregation period is regarded as 0, that is, each participant is not in the current Contribute during the aggregation period. The following is a detailed description of the process of using a calculator to determine whether to calculate the contribution value of each participant in this aggregation cycle in combination with a specific embodiment, which may specifically include the following:

If |v _N -v ₀ |≤λ, add 1 to the counter, if the counter exceeds 1 (the value of the counter is greater than or equal to 2), then the contribution value of each participant i in this aggregation period t

Then return to the previous step, otherwise continue to the next step; in other words, by making a difference between the final utility value corresponding to the aggregation model generated before and after the current aggregation cycle and the initial utility value, and comparing the difference with the first cut-off threshold, When the difference is less than the first truncation threshold, add 1 to the value of the counter. If the counter is greater than or equal to 2 (that is, the utility change value of the aggregation model generated by two consecutive rounds of aggregation cycles is less than the threshold value), then it is judged that within the aggregation cycle The contribution value of each participant is 0.

The purpose of this embodiment of the present disclosure is to evaluate the changes in the utility value of the joint model in the current round before formally calculating the contribution value of each participant. Whether the utility value of the joint model has been improved, that is, whether the performance of the joint model itself has been improved. If the improvement in model performance is small, it can be considered that the contribution value of each participant in this round is 0; The performance improvement of the learning model is relatively large, so continue to perform the following calculations, that is, specifically calculate the contribution value of each participant in this round. This disclosure can judge in advance whether it is necessary to further calculate the contribution value of the participant, or directly count the contribution value of the participant in the current round as 0, thereby avoiding an invalid calculation process and improving calculation efficiency.

In some embodiments, a participant combination is selected from all participant combinations in a predetermined order, and the marginal contribution value of each participant in the selected participant combination is estimated, including: The arrangement order of the combination of participants is selected in turn, and the participants in the selected combination of participants are iterated sequentially, so as to estimate the marginal contribution value generated when the participant joins the combination of participants, and obtain the marginal contribution An estimate of the value.

Specifically, when it is judged that the contribution value of each participant in the current round needs to be calculated, a participant combination P is sequentially selected from all participant combinations Ps in the current round, and the selection order and enumeration here are all possible The order of the participant combinations Ps is the same, that is, the combination with a small number of participants is selected first, and then the combination with a large number of participants is selected progressively.

Further, in the process of sequentially selecting each participant combination P, for each participant combination P, estimate the marginal contribution value of each participant in the participant combination, that is, calculate the The marginal contribution value generated during the combination; below in conjunction with a specific embodiment, the estimation process of the marginal contribution value of each participant in the participant combination P will be described in detail, which may specifically include the following:

For each participant j in P, P can be divided into two subsets {j} and S=P\{j}, that is, P=S∪{j}; calculate the marginal contribution generated by adding j to S, in When calculating the marginal contribution, since the actual marginal contribution should be Δ _{j_real} =v _S∪{j} -v _S =V(S∪{j})-V(S)=V(P)-V(S); but , the value of V(P) cannot be determined temporarily, so the scaling principle is used to enlarge V(P) to v _N , so that the marginal contribution of j is estimated as Δ _{j_est} =v _N -v _S =v_lut[N]-v_lut [S], and add v _S to the list VS_hist (corresponding to another cache table), since v _S has been calculated in the previous combination P', here only needs to be obtained from the lookup table v_lut, no need to calculate V(M _S ^(t) ).

Further, use the above calculation method to perform a cycle for each element (participant j) in the participant combination P, until each element in the participant combination P is cyclically calculated once, and finally estimate that in each participant combination The marginal contribution value corresponding to each participant.

In some embodiments, according to the estimation result and the weight of the participant combination, it is judged whether to use an interpolation function to calculate the utility value of the participant combination, including: calculating the estimated value of the marginal contribution value corresponding to the participant and the corresponding participant Combined weight product, and compare the product with the preset second truncation threshold; when the product corresponding to each participant in the participant combination is less than or equal to the preset second truncation threshold, it is judged to use the interpolation function Calculate the utility value of the participant combination; otherwise, use the preset model deduction method to calculate the utility value of the participant combination.

Specifically, according to the relationship between the product of the marginal contribution value of the participant and the weight of the participant combination corresponding to the participant and the second cut-off threshold, it is judged whether to calculate the utility value of the participant combination; the following is combined with specific embodiments , to describe the calculation and judgment process of the above product in detail, which may specifically include the following:

_Calculate the product |Δ _{j_est} *w _| _S|| _|| ≤η*|vN-v0|, then judge to use the interpolation function to calculate the utility value of the participant combination P, otherwise, judge to use the preset model deduction method to calculate the utility value of the participant combination P.

In other words, if the marginal contribution value of each participant in the participant combination P satisfies the above formula, then the utility value of the participant combination P does not need to be deduced at this time, and the interpolation function is directly used to calculate the participant combination The utility value of P, if the marginal contribution value of one participant is not satisfied, it is necessary to perform model deduction on the participant combination P and calculate the utility value.

It should be noted that when the marginal contribution value of each participant in the above participant combination P satisfies the above formula, it is only an optional embodiment. In addition, other judgment criteria can also be set, such as The marginal contribution value of any participant in the participant combination P does not satisfy the above formula, or half of the participants in the participant combination P do not satisfy the above formula, or there is any proportion of participants in the participant combination P that does not satisfy the above formula formulas and more.

According to the technical solution provided by the embodiments of the present disclosure, in order to judge in advance whether to perform model derivation on the participant combination, the marginal contribution value of each participant in the participant combination is estimated by amplifying the utility value, and the estimated value Multiply with the weight of the participant combination, and compare the product with the second cut-off threshold, so as to judge whether to use the interpolation function to calculate the utility value of the participant combination, or to use the model deduction to calculate the utility value of the participant combination; because the model The complexity of the deduction is very high, and the amount of calculation is large. Therefore, by adding the above-mentioned judgment means, for the combination of participants that does not need to do model deduction, it can be directly obtained by weighting and summing the utility values of the sub-combinations calculated in the previous iteration process. Thus, the calculation speed of the contribution value is improved.

In some embodiments, using an interpolation function to calculate the utility value of the participant combination includes: based on the utility value of the participant combination calculated in the historical iteration process, and the corresponding utility value when the participant combination is a complete set of participant combinations, using a preset The set interpolation function estimates the utility value of the participant combination, obtains the estimated value corresponding to the utility value of the participant combination, and updates the lookup table according to the estimated value.

Specifically, the principle of the interpolation function to calculate the utility value is to use the calculated utility value of the subcombination to estimate the utility value of the participant combination, that is, to calculate the utility value of the participant combination according to the utility value of the subcombination in the previous iteration process. Approximation processing, for example, for the participant combination (1, 2, 3), the values of v(1, 2), v(2, 3), v(1, 3) can be used to estimate v(1, 2, 3), and the value of v(1, 2),, v(2, 3), v(1, 3) already exists in the v_lut lookup table.

Further, the calculation formula of the interpolation function V(P)=interpolate(mean(VS_hist),v _N ,P,N), according to the calculation formula, obtain the utility estimate of the participant combination P and update the lookup table v_lut[P]= V(P). Among them, the implementation method of the interpolation function interpolate( ) is:

In the calculation formula of the above interpolation function, mean(VS_hist), v _N , sub-combination P and full combination N are used as the input of the function, calculated and output.

Furthermore, in the calculation formula of the interpolation function above, the method of calculating the average value (that is, the mean in the formula) is used to calculate the estimated utility value of the participant combination P, but in practical applications, in addition to the method of calculating the average value, The utility estimation value of the participant combination P can also be calculated by calculating the maximum value, the minimum value or the median value. When the maximum value method is used for calculation, the mean in the above formula can be directly replaced by max; when the minimum value method is used for calculation, the mean in the above formula can be directly replaced by min. Therefore, the calculation method of the interpolation function does not constitute a Limitations on the technical solution of the application.

In some embodiments, the utility value of the participant combination is calculated using a preset model derivation method, and the lookup table is updated according to the calculated utility value of the participant combination, including: aggregating the model parameters corresponding to the participant combination , and perform model deduction on the model corresponding to the participant combination, aggregate the weights of each participant in the participant combination to obtain the weight of the participant combination, perform model deduction on the standard verification set for the participant combination, and calculate the participation The real utility value of the square combination is used to update the lookup table.

Specifically, the model deduction adopts the formula V(P)=V(M _P ^(t) )=V(Agg(P)), according to this calculation formula, the real utility value of the participant combination P is obtained and the lookup table v_lut[P] is updated =V(P).

In some embodiments, calculating the contribution value of each participant based on the final updated lookup table includes: using the preset Shapley The value calculation formula calculates the contribution value corresponding to each participant, where the contribution value is used to represent the contribution of the participant to the joint learning model trained in the aggregation cycle in the joint learning.

Specifically, according to the calculation process of the utility value of the participant combination P in the above embodiment, the entire participant combination Ps is cycled once to obtain the final updated v_lut lookup table, and the corresponding utility of the participant combination in the lookup table is obtained. Values are substituted into the preset Shapley value calculation formula in turn to calculate the contribution value corresponding to each participant, and then the contribution value of each contributor in the current aggregation period can be obtained.

Furthermore, Shapley Value is a method for fairly distributing benefits based on the average marginal contribution of individual i joining combination S, and its computational complexity is O(2 ⁿ ), where n is the total number of individuals. Its calculation formula is:

The Shapley value (that is, the Shapley value) considers all possible orders in which individual i joins the sub-combination, where N represents the full combination, S represents the sub-combination in a certain arrangement, V(·) represents the utility function, and the |·| symbol Indicates the number of elements in the set, [V(S∪{i})–V(S)] indicates the marginal utility of i added to the sub-combination S, and the weight w _|S| = |S|! (|N|-|S|-1)! /|N|! Indicates the probability of the combination occurring.

Further, repeat the above steps to obtain the contribution value of each participant i in all T aggregation periods, and accumulate the contribution value of participant i to the joint model; that is, calculate all the aggregation periods according to the above method, and obtain each participant i Party’s corresponding contribution value in each aggregation cycle, and then accumulate to get the total contribution value.

Further, the first truncation threshold λ is set in the following manner: Let the marginal gain of the final joint model utility function relative to the initial model be Δ _U =|V(M ^(T) )-V(M ⁽⁰⁾ )| , where T is the total communication round, which can be set to λ=Δ _U *0.01. The second truncation threshold η can be used to represent the error level of the contribution value, which can be set as η=1e-3˜1e-5 _.

The following are device embodiments of the present disclosure, which can be used to implement the method embodiments of the present disclosure. For details not disclosed in the disclosed device embodiments, please refer to the disclosed method embodiments.

Fig. 4 is a schematic structural diagram of an apparatus for determining a participant's contribution in joint learning provided by an embodiment of the present disclosure. As shown in Figure 4, the means for determining the contribution of participants in this joint learning include:

The construction module 401 is configured to determine the participants of the joint learning, construct all participant combinations based on the participants, and calculate the weight corresponding to each participant combination;

The first judging module 402 is configured to determine the utility change value corresponding to the joint learning model before and after the aggregation period and establish a lookup table, and judge whether to calculate the contribution value of the participant in the aggregation period according to the utility change value;

The second judging module 403 is configured to, when the judging result is yes, select a participant combination from all participant combinations in a predetermined order, and estimate the marginal contribution value of each participant in the selected participant combination , according to the estimation result and the weight of the participant combination, judge whether to use the interpolation function to calculate the utility value of the participant combination;

The update module 404 is configured to use an interpolation function to calculate the utility value of the participant combination when the judgment result is yes, and to calculate the utility value of the participant combination by using a preset model derivation method when the judgment result is no, and calculate according to Update the lookup table with the utility value of the obtained participant combination;

The calculation module 405 is configured to select each participant combination in turn until the utility values of all participant combinations are calculated, and update the lookup table by using the utility values of all participant combinations to obtain a final updated lookup table, so that based on The final updated lookup table calculates the contribution value of each participant, and determines the contribution of the participant in the joint learning according to the contribution value.

In some embodiments, the construction module 401 in FIG. 4 constructs all participant combinations according to the number of participants from the fewest to the largest according to all participants in the joint learning. All participant combinations include multiple participant combinations. According to The number of participants in each participant combination calculates the weight; wherein, the weight is used to represent the probability that the participant combination appears in all the participant combinations.

In some embodiments, the first judgment module 402 in FIG. 4 determines the initial utility value and the final utility value of the aggregation period for each aggregation period, and uses the difference between the final utility value and the initial utility value as the utility change value, And establish a lookup table containing all participant combinations corresponding to the aggregation cycle; perform an initialization operation on the lookup table so that the initial utility values of other participant combinations in the lookup table except the empty set participant combination and the full set participant combination are 0; among them, the lookup table is used to store the utility values corresponding to the combinations of all participants.

In some embodiments, the first judging module 402 in FIG. 4 compares the utility change value of the aggregation period with the preset first cut-off threshold, and when the utility change value of the aggregation period is less than the first cut-off threshold, and multiple rounds of aggregation When the utility change values of the period are all less than the first cut-off threshold, it is judged that the contribution value of each participant in the aggregation period is 0; otherwise, the contribution value of each participant in the aggregation period is calculated.

In some embodiments, the second judging module 403 in FIG. 4 sequentially selects the participant combinations according to the arrangement order of the participant combinations in all the participant combinations, and iterates the participants in the selected participant combinations sequentially. , in order to estimate the marginal contribution value generated when the participant joins the participant combination, and obtain the estimated value of the marginal contribution value.

In some embodiments, the second judging module 403 in FIG. 4 calculates the product of the estimated value of the participant's corresponding marginal contribution value and the weight of the participant's corresponding participant combination, and compares the product with the preset second cut-off threshold ; When the product corresponding to each participant in the participant combination is less than or equal to the preset second cut-off threshold, it is judged to use the interpolation function to calculate the utility value of the participant combination; otherwise, use the preset model derivation method Computes the utility value for a combination of parties.

In some embodiments, the update module 404 in FIG. 4 uses a preset interpolation function to update the participant's The utility value of the combination is estimated, and the estimated value corresponding to the utility value of the participant combination is obtained, and the lookup table is updated according to the estimated value.

In some embodiments, the update module 404 in FIG. 4 aggregates the model parameters corresponding to the participant combination, performs model deduction on the model corresponding to the participant combination, and aggregates the weight of each participant in the participant combination to obtain The weight of the participant combination, the model deduction of the participant combination on the standard verification set is calculated to obtain the real utility value of the participant combination, and the lookup table is updated using the real utility value.

In some embodiments, the calculation module 405 in FIG. 4 calculates the contribution value corresponding to each participant by using the preset Shapley value calculation formula according to the utility value of each participant combination in the finally updated lookup table , where the contribution value is used to represent the contribution of the participants to the joint learning model trained in the aggregation cycle in the joint learning.

It should be understood that the sequence numbers of the steps in the above embodiments do not mean the order of execution, and the execution order of each process should be determined by its functions and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present disclosure.

Fig. 5 is a schematic flowchart of another method for determining the contribution of a participant in joint learning provided by an embodiment of the present disclosure. The method for determining the contribution degree of a participant in the joint learning in FIG. 2 may be executed by a server of the joint learning. As shown in Figure 2, the method for determining the contribution of the participants in the joint learning may specifically include:

S501. Determine all participant combinations based on the participants in the joint learning, and calculate the weight of each participant combination in all participant combinations;

S502. Determine the first utility value of the joint model before the start of the current aggregation period, and the second utility value of the joint model after the end of the current aggregation period, calculate the utility change value based on the first utility value and the second utility value, and establish a search Table; among them, the utility change value is used to judge whether to calculate the contribution value of each participant in the current aggregation period;

S503, when judging and calculating the contribution value, select a participant combination from all the participant combinations, and calculate the marginal contribution value corresponding to each participant in the participant combination, and judge to use the first estimation method according to the marginal contribution value and weight Or the second estimation method estimates the utility value of the participant combination;

S504. Determine the estimation result of the utility value of the participant combination by the first estimation method or the second estimation method, use the estimation result to update the lookup table, and iteratively estimate the utility value of each participant combination in turn, and obtain the basis utility value The final lookup table obtained after the update is used to calculate the contribution value of the participant using the final lookup table.

According to the technical solution provided by the embodiments of the present disclosure, all participant combinations are determined based on the participants in the joint learning, and the weight of each participant combination in all participant combinations is calculated; the first joint model of the joint model before the current aggregation cycle starts A utility value, and the second utility value of the joint model after the end of the current aggregation period, calculate the utility change value based on the first utility value and the second utility value, and establish a lookup table; wherein, the utility change value is used to judge whether to calculate the current The contribution value of each participant in the aggregation period; when judging and calculating the contribution value, select a participant combination from all participant combinations, and calculate the marginal contribution value corresponding to each participant in the participant combination, according to the marginal contribution value and Weight, to judge the utility value of the participant combination by using the first estimation method or the second estimation method; determine the estimation result of the utility value of the participant combination by the first estimation method or the second estimation method, and use the estimation result to look up the table Update, and iteratively estimate the utility value of each participant combination in turn, and obtain the final lookup table obtained after updating according to the utility value, so as to use the final lookup table to calculate the contribution value of the participants. The present disclosure can improve the calculation accuracy of the contribution value in joint learning, reduce the amount of calculation data, and make the calculation result of the contribution value more accurate and the calculation efficiency higher.

The cyclic process of calculating the contribution value of each participant in the joint learning of the present disclosure will be described in detail below in combination with a specific schematic diagram of the program flow. FIG. As shown in Figure 6, the program for calculating the contribution value of the participants may specifically include the following:

In some embodiments, determining all participant combinations based on the participants in the joint learning, and calculating the weight of each participant combination in all the participant combinations includes: determining the participants in the joint learning, dividing the participants according to the number from Enumerate multiple participant combinations sequentially in the least-most way, and use the set of multiple participant combinations as all participant combinations; calculate the weight corresponding to the participant combination based on the number of participants in the participant combination, where, The weight is used to represent the probability that the participant combination appears in all the participant combinations.

Assuming that there are

N participants

1, 2, ... i...n-1, n joint learning, the training has been aggregated for T cycles, and each aggregation cycle t in the training process is recorded, and the local data uploaded by each participant i Model M _i ^(t) , and the joint model M ^(t) after central aggregation, initialize the model M ⁽⁰⁾ , have an evaluation function or utility function V( ) for model performance (such as accuracy, loss, etc.), and the joint Model aggregation method Agg(·), thresholds λ, η; where λ represents the first truncation threshold, and η represents the second truncation threshold.

Further, firstly, according to all participants in the joint learning, enumerate all possible participant combinations Ps=[(1,),(2,),(3,)...,(1 ,2),(1,3),(2,3),…P,…N]; for each subcombination S with 0, 1, 2,…n-1 participants, calculate the weight w _{|S |} = |S|! (|N|-|S|-1)! /|N|! .

In some embodiments, determine the first utility value of the joint model before the start of the current aggregation period, and the second utility value of the joint model after the end of the current aggregation period, and calculate the utility change value based on the first utility value and the second utility value , and establish a lookup table, including: determine the first utility value and the second utility value corresponding to the joint model before and after the current aggregation period, calculate the difference between the second utility value and the first utility value, and convert the difference value as the utility change value, and establish a lookup table containing all participant combinations corresponding to the current aggregation cycle; perform an initialization operation on the lookup table so that all participants in the lookup table except the empty set participant combination and the full set participant combination The initial utility value of the combination of parties is 0; wherein, the lookup table is used to store the corresponding utility values of all combinations of participants.

Specifically, calculate the utility value corresponding to the joint model at the beginning and after the end of each aggregation period, and establish a lookup table, that is, for each aggregation period, you can first calculate the final utility of this aggregation period (that is, the current aggregation period) value and the initial utility value of this aggregation period, the calculation process of the initial utility value and the final utility value of the aggregation period will be described in detail in combination with a specific embodiment, which may specifically include the following:

In some embodiments, the utility change value is used to determine whether to calculate the contribution value of each participant in the current aggregation period, including: comparing the utility change value of the joint model corresponding to the current aggregation period with the preset first cut-off threshold, when When the utility change value is less than the first truncation threshold, it is judged that the contribution value of each participant in the current aggregation period is 0; otherwise, the contribution value of each participant in the current aggregation period is recalculated.

Specifically, by calculating the final utility value of this aggregation period and the initial utility value of this aggregation period, if the difference between the final utility value corresponding to this aggregation period and the initial utility value is less than the first cut-off threshold, the calculation ends, and The contribution value of each participant in the current aggregation period is regarded as 0, that is, each participant does not make a contribution in the current aggregation period. The following is a detailed description of the process of judging whether to calculate the contribution value of each participant in this aggregation cycle in combination with a specific embodiment, which may specifically include the following:

In the current aggregation period, if it is judged that |v _N -v ₀ |≤λ, then the contribution value of each participant i in this aggregation period t

And return to the previous step, otherwise continue to the next step; in other words, by making a difference between the final utility value corresponding to the aggregation model generated before and after the current aggregation cycle and the initial utility value, and comparing the difference with the first cut-off threshold, When the difference is smaller than the first truncation threshold, the contribution value of each participant in this aggregation period is directly judged as 0.

The purpose of this embodiment of the present disclosure is to evaluate the changes in the utility value of the joint model in the current round before formally calculating the contribution value of each participant. Whether the utility value of the joint model has been improved, that is, whether the performance of the joint model itself has been improved. If the improvement in model performance is small, it can be considered that the contribution value of each participant in this round is 0; The performance improvement of the model is relatively large, so continue to perform the following calculations, that is, specifically calculate the contribution value of each participant in this round. This disclosure can judge in advance whether it is necessary to further calculate the contribution value of the participant, or directly count the contribution value of the participant in the current round as 0, thereby avoiding an invalid calculation process and improving calculation efficiency.

In some embodiments, a participant combination is selected from all participant combinations, and the marginal contribution value corresponding to each participant in the participant combination is calculated, including: according to the arrangement order of each participant combination in all participant combinations , select a participant combination from all participant combinations in turn, and randomly select a participant from the participant combination; divide a sub-combination from the participant combination according to the participant, and calculate the margin generated when the participant joins the sub-combination Contribution value, iterates the participants in the participant combination in order to calculate the marginal contribution value corresponding to each participant; where, the sub-combination is a set of participants in the participant combination except randomly selected participants .

Further, use the above calculation method to perform a cycle for each element in the participant combination P (participant i) until each element in the participant combination P is calculated in a cycle, and finally estimated to obtain The marginal contribution value corresponding to each participant.

In some embodiments, according to the marginal contribution value and the weight, it is judged to use the first estimation method or the second estimation method to estimate the utility value of the participant combination, including: calculating the marginal contribution value of the participant and the participant combination to which the participant belongs The product between the weights, and compare the product with the preset second truncation threshold; when the product corresponding to each participant in the participant combination is less than or equal to the second truncation threshold, it is judged to use the first estimation method Estimate the utility value of the participant combination; otherwise, use the second estimation method to estimate the utility value of the participant combination.

Calculate the product | _Δ _{j_est} *w _| _S|| _|| ≤η*|vN-v0|, it is judged to use the utility value of the subcombination to estimate the utility value of the participant combination P, otherwise, it is judged to use the preset model derivation method to calculate the utility value of the participant combination P.

That is to say, if the marginal contribution value of each participant in the participant combination P satisfies the above formula, then the utility value of the participant combination P does not need to be deduced at this time, and the utility value of the subcombination can be directly used to determine Estimate the utility value of the participant combination P. If the marginal contribution value of one participant is not satisfied, it is necessary to model the participant combination P and calculate the utility value.

According to the technical solution provided by the embodiments of the present disclosure, in order to judge in advance whether to perform model derivation on the participant combination, the marginal contribution value of each participant in the participant combination is estimated by amplifying the utility value, and the estimated value Multiply with the weight of the participant combination, and compare the product with the second cut-off threshold, so as to judge whether to use the utility value of the sub-combination to estimate the utility value of the participant combination, or to calculate the utility value of the participant combination by model deduction; Due to the high complexity of model deduction and the large amount of calculation, by adding the above-mentioned judgment means, for the combination of participants that do not need to do model deduction, the utility value of the sub-combination that has been calculated in the previous iteration process can be directly used to estimate The utility value of the participant combination, thus improving the calculation speed of the contribution value.

In some embodiments, using the first estimation method to estimate the utility value of the participant combination includes: obtaining the utility value corresponding to the sub-combination in the participant combination from a lookup table, and averaging the utility values corresponding to the sub-combination , the calculation of the maximum value or minimum value, using the calculated average value, maximum value or minimum value as the estimated value of the utility value of the combination of participants, and updating the lookup table according to the estimated value.

Specifically, the first estimation method is to use the utility value corresponding to the sub-combination in the participant combination to estimate the utility value of the participant combination. When using the calculated utility value of the sub-combination to estimate the utility value of the participant combination , by calculating the average, maximum or minimum value among the utility values of the subcombinations, so as to approximate the utility value of the participant combination, for example, for the participant combination (1, 2, 3), The value of v(1,2,3) can be estimated by the average, maximum or minimum value of v(1,2), v(2,3), v(1,3), and v(1, 2), the values of v(2, 3), v(1, 3) already exist in the v_lut lookup table.

Further, taking the calculation of the average value between the utility values of the sub-combinations as an example, the calculation process of the utility value of the participant combination is described in detail, which may specifically include the following formula:

V(P)=mean(VS_hist)

Among them, V(P) represents the utility value of the participant combination, and VS_hist represents the utility value corresponding to the sub-combination; according to the average calculation formula, the utility estimate of the participant combination P can be obtained and the lookup table v_lut[P]=V can be updated (P).

Furthermore, in the above calculation formula, the method of calculating the mean value (i.e. the mean in the formula) is used to calculate the utility estimate of the participant combination P, but in practical applications, in addition to the method of calculating the mean value, there is also The utility estimation value of the participant combination P can be calculated by calculating the maximum value or the minimum value. When the maximum value method is used for calculation, the mean in the above formula can be directly replaced by max; when the minimum value method is used for calculation, the mean in the above formula can be directly replaced by min. Therefore, the calculation method of the above utility value does not constitute Restrictions on the technical solution of this application.

In some embodiments, using the second estimation method to estimate the utility value of the participant combination includes: aggregating the model parameters corresponding to the participant combination, and performing model deduction on the model corresponding to the participant combination, and calculating the participant combination The weight of each participant is aggregated to obtain the weight of the participant combination, and the model deduction is carried out on the standard verification set for the participant combination, and the real utility value of the participant combination is calculated, and the lookup table is updated using the real utility value.

Specifically, the second estimation method is to use the preset model derivation method to calculate the utility value of the participant combination, and the model deduction adopts the formula V(P)=V(M _P ^(t) )=V(Agg(P)), according to The calculation formula obtains the real utility value of the participant combination P and updates the lookup table v_lut[P]=V(P).

In some embodiments, obtaining the final lookup table obtained after updating according to the utility value, so as to use the final lookup table to calculate the contribution value of the participant, includes: obtaining the utility value of the participant combination in the final lookup table, using the preset The Shapley value formula is used to calculate the contribution value corresponding to each participant, where the contribution value is used to represent the contribution value of the participant to the joint model trained in the aggregation cycle in the joint learning.

The Shapley value (that is, the Shapley value) considers all possible orders in which individual i joins the sub-combination, where N represents the full combination, S represents the sub-combination in a certain arrangement, V(·) represents the utility function _, |·| symbol Indicates the number of elements in the set, [V(S∪{i})–V(S)] indicates the marginal utility of i added to the sub-combination S, and the weight w _|S| = |S|! (|N|-|S|-1)! /|N|! Indicates the probability of the combination occurring.

Fig. 7 is a schematic structural diagram of another apparatus for determining a participant's contribution in joint learning provided by an embodiment of the present disclosure. As shown in Figure 7, the means for determining the contribution of participants in this joint learning include:

The determining module 701 is configured to determine all participant combinations based on the participants in the joint learning, and calculate the weight of each participant combination in all participant combinations;

The establishment module 702 is configured to determine a first utility value of the joint model before the start of the current aggregation period, and a second utility value of the joint model after the end of the current aggregation period, and calculate a utility change based on the first utility value and the second utility value value, and establish a lookup table; among them, the utility change value is used to judge whether to calculate the contribution value of each participant in the current aggregation period;

The judging module 703 is configured to select a participant combination from all participant combinations when judging and calculating the contribution value, and calculate the marginal contribution value corresponding to each participant in the participant combination, and judge according to the marginal contribution value and weight Estimate the utility value of the participant combination by using the first estimation method or the second estimation method;

The calculation module 704 is configured to determine the estimation result of the utility value of the participant combination by the first estimation method or the second estimation method, use the estimation result to update the lookup table, and iteratively estimate to obtain the utility value of each participant combination , to obtain the final lookup table obtained after updating according to the utility value, so as to use the final lookup table to calculate the contribution value of the participant.

In some embodiments, the determination module 701 in FIG. 7 determines the participants in the joint learning, and enumerates the participants in order from the least to the largest to obtain a combination of multiple participants. The combination of multiple participants constitutes The set is used as a combination of all participants; the weight corresponding to the combination of participants is calculated based on the number of participants in the combination of participants, where the weight is used to represent the probability that the combination of participants appears in all combinations of participants.

In some embodiments, the establishment module 702 in FIG. 7 determines the first utility value and the second utility value corresponding to the joint model before and after the current aggregation period, and calculates the difference between the second utility value and the first utility value Value, use the difference as the utility change value, and establish a lookup table containing all participant combinations corresponding to the current aggregation cycle; perform an initialization operation on the lookup table, so that the lookup table excludes the empty set participant combination and the full set participant combination The initial utility value of other participant combinations other than 0; wherein, the lookup table is used to store the utility values corresponding to all participant combinations.

In some embodiments, the establishment module 702 in FIG. 7 compares the utility change value of the joint model corresponding to the current aggregation period with the preset first cut-off threshold, and when the utility change value is less than the first cut-off threshold, then judge the current aggregation period The contribution value of each participant in the period is 0; otherwise, recalculate the contribution value of each participant in the current aggregation period.

In some embodiments, the judging module 703 in FIG. 7 sequentially selects a participant combination from all participant combinations according to the arrangement order of each participant combination in all participant combinations, and randomly selects a participant combination from the participant combination. According to the participant's division of a sub-combination from the participant combination, the marginal contribution value generated when the participant joins the sub-combination is calculated, and the participants in the participant combination are iterated in order to calculate the corresponding marginal contribution value; where, the sub-combination is a set of participants in the participant combination except randomly selected participants.

In some embodiments, the judging module 703 in FIG. 7 calculates the product of the marginal contribution value of the participant and the weight of the participant group to which the participant belongs, and compares the product with the preset second cut-off threshold; when the participant When the product corresponding to each participant in the combination is less than or equal to the second cut-off threshold, it is judged to use the first estimation method to estimate the utility value of the participant combination; otherwise, use the second estimation method to estimate the utility value of the participant combination Make an estimate.

In some embodiments, the judging module 703 in FIG. 7 obtains the utility value corresponding to the sub-combination in the participant combination from the lookup table, calculates the average value, maximum value or minimum value of the utility value corresponding to the sub-combination, and calculates The obtained average value, maximum value or minimum value is used as an estimated value of the utility value of the participant combination, and the lookup table is updated according to the estimated value.

In some embodiments, the judging module 703 in FIG. 7 aggregates the model parameters corresponding to the participant combination, performs model deduction on the model corresponding to the participant combination, and aggregates the weight of each participant in the participant combination to obtain The weight of the participant combination, the model deduction of the participant combination on the standard verification set is calculated to obtain the real utility value of the participant combination, and the lookup table is updated using the real utility value.

In some embodiments, the calculation module 704 in FIG. 7 obtains the utility value of the combination of participants in the final lookup table, and uses the preset Shapley value formula to calculate the contribution value corresponding to each participant, wherein the contribution value is used Yu represents the contribution value of the participant to the joint model trained in the aggregation cycle in the joint learning.

Fig. 8 is a schematic flowchart of another method for determining the contribution degree of a participant in joint learning provided by an embodiment of the present disclosure. The method for determining the contribution degree of a participant in the joint learning of FIG. 8 may be executed by a server of the joint learning. As shown in Figure 8, the method for determining the contribution of the participants in the joint learning may specifically include:

S801. Based on the framework of joint learning, generate multiple participant groups, determine a participant group set composed of multiple participant groups, and calculate the weights of the participant groups, wherein each participant group includes at least two participants;

S802. Determine the aggregation period in the joint learning, obtain the utility change value corresponding to the joint learning model before and after the aggregation period and establish a lookup table, and judge whether to calculate the contribution value of each participant in the aggregation period according to the utility change value;

S803, when the judgment result is yes, use the participant groups in the participant group set to randomly generate a full permutation combination, and generate a plurality of subcombinations according to the order of the participants in the participant group in the full permutation combination, and calculate the participant joining subcombination The estimated value of the marginal contribution value at the time of combination, according to the estimated value of the marginal contribution value and the weight of the participant group, judge whether to use an interpolation function to calculate the utility value of the new participant group formed after the participant joins the sub-combination;

S804, when the judgment result is yes, use the interpolation function to calculate the utility value of the new participant group; when the judgment result is no, use the preset model deduction method to calculate the utility value of the new participant group, and calculate The utility value of the participant group updates the lookup table;

S805, based on the updated lookup table, calculate the marginal contribution value of the participant, and judge whether the marginal contribution value of the participant is converged, and when the judgment result is yes, use the converged marginal contribution value as the contribution value of the participant, when When the judgment result is no, a new full permutation combination is generated until the contribution values of all converged participants are calculated, and the contribution of the participants in the joint learning is determined according to the contribution values.

According to the technical solution provided by the embodiments of the present disclosure, multiple participant groups are generated through a framework based on joint learning, and a set of participant groups composed of multiple participant groups is determined to calculate the weights of the participant groups, where each participant The square group contains at least two participants; determine the aggregation period in the joint learning, obtain the utility change value corresponding to the joint learning model before and after the aggregation period and establish a lookup table, and judge whether to calculate the contribution value of each participant in the aggregation period according to the utility change value ; When the judgment result is yes, use the participant groups in the participant group set to randomly generate a full permutation combination, and generate multiple subcombinations according to the order of the participants in the participant group in the full permutation combination, and calculate the participants to join the subcombination According to the estimated value of the marginal contribution value and the weight of the participant group, it is judged whether to use the interpolation function to calculate the utility value of the new participant group formed after the participant joins the sub-combination; when the judgment result When it is yes, use the interpolation function to calculate the utility value of the new participant group; when the judgment result is no, use the preset model deduction method to calculate the utility value of the new participant group, and based on the calculated utility of the new participant group value to update the lookup table; based on the updated lookup table, calculate the marginal contribution value of the participant, and judge whether the marginal contribution value of the participant is converged, and when the judgment result is yes, use the converged marginal contribution value as the participant When the judgment result is no, a new full permutation combination is generated until the contribution values of all converged participants are calculated, and the contribution of the participants in the joint learning is determined according to the contribution values. The present disclosure can improve the calculation accuracy of the contribution value in joint learning, reduce the amount of calculation data, and make the calculation result of the contribution value more accurate and the calculation efficiency higher.

The cyclic process of calculating the contribution value of each participant in the joint learning of the present disclosure will be described in detail below in combination with a specific schematic diagram of the program flow. FIG. As shown in Figure 9, the program for calculating the contribution value of the participants may specifically include the following:

In some embodiments, generating a plurality of subcombinations according to the order of the participants in the participant group in the full permutation combination, and calculating the estimated value of the contribution margin when the participant joins the subcombination includes: dividing the full permutation combination into Multiple sub-combinations, and determine the next participant corresponding to the last participant of the sub-combination in the full permutation combination, and calculate the estimated value of the marginal contribution value of the next participant when joining the sub-combination.

Specifically, when it is judged that it is necessary to calculate the contribution value of each participant in the current round, first randomly generate a full permutation combination P from the participant group set Ps in the current round of aggregation period, and set k=0, for example: suppose A participant group set Ps contains

elements

1, 2, 3, 4, and 5. These 5 elements correspond to 5 participants respectively. According to these 5 participants, the following full permutations and combinations can be generated: (1, 2, 3, 4, 5), (2, 3, 4, 5, 1), (3, 4, 5, 1, 2), (4, 5, 1, 2, 3), (5, 4, 3, 2, 1)……. After randomly generating a full permutation combination, the participants in the full permutation combination are divided into multiple sub-combinations.

Further, when calculating the marginal contribution value of a participant based on the full permutation combination, the following methods can be used:

The first j participants are sequentially taken from the full permutation combination P to form a subcombination S, and an estimated value of the marginal contribution generated when the j+1th participant joins the subcombination S is calculated. For example, the full permutation combination is (5, 4 . An estimate of the contribution margin. In practical applications, the following formula can be used to estimate the marginal contribution, that is, Δ _{j+1_est} =v _N -v _S =v_lut[N]-v_lut[S], since v _S has been calculated in the previous combination S' , here only needs to be obtained from the lookup table v_lut, no need to calculate V(M _S ^(t) ).

Further, use the above calculation method to perform a loop on all the sub-combinations S formed in the current full permutation combination, and the marginal contribution value generated when the j+1th participant joins the sub-combination S, until the current full permutation combination The situation of each sub-combination is calculated once, and finally the marginal contribution value of the participant when joining the sub-combination is estimated.

In some embodiments, according to the estimated value of the marginal contribution value and the weight of the participant group, it is judged whether to use an interpolation function to calculate the utility value of the new participant group formed after the participant joins the sub-combination, including: calculating the marginal contribution value The product of the estimated value and the weight of the participant group, and compare the product with the preset second cut-off threshold; when the products are all less than or equal to the second cut-off threshold, it is judged to use the interpolation function to calculate the utility of the new participant group value; otherwise, use the preset model derivation method to calculate the utility value of the new participant group.

Specifically, according to the relationship between the product of the estimated value of the marginal contribution value and the weight of the participant group and the second cut-off threshold, it is judged whether to calculate the utility value of the new participant group; The calculation and judgment process will be described in detail, which may include the following:

Calculate the product |Δ _{j+1_est} *w _|S| | of the marginal contribution value generated when the j+1th participant joins the sub-combination S and the weight corresponding to the sub-combination S, if the product satisfies |Δ _{j+1_est} * ^w _|S| |≤η*|vN-v0|, then judge to use the interpolation function to calculate the utility value of the new participant group, otherwise, judge to use the preset model derivation method to calculate the utility value of the new participant group.

That is to say, if the marginal contribution value generated when the j+1th participant joins the subgroup S satisfies the above formula, then the utility value corresponding to the new participant group after the j+1th participant joins the subgroup S There is no need to perform deduction, and the interpolation function is directly used to calculate the utility value of the new participant group. If the above formula is not satisfied, it is necessary to perform model deduction for the new participant group and calculate the utility value.

According to the technical solution provided by the embodiments of the present disclosure, in order to judge in advance whether to perform model deduction on the new participant group, the marginal contribution value generated when the j+1th participant joins the subgroup S is estimated by amplifying the utility value , and multiply the estimated value with the weight of the participant group, and compare the product with the second cut-off threshold, so as to determine whether to use the interpolation function to calculate the utility value of the new participant group or to use model deduction to calculate the utility value of the new participant group The utility value of the group; due to the high complexity of model deduction and the large amount of calculation, by adding the above-mentioned judgment means, for the new participant group that does not need to do model deduction, the sub-combination calculated in the previous iteration process can be directly used. The utility value is obtained by weighted summation, which improves the calculation speed of the contribution value.

In some embodiments, using an interpolation function to calculate the utility value of the new participant group includes: based on the utility value of the sub-combination calculated in the historical iteration process, and the corresponding utility value when the participant group is a full set of participant groups, using a preset The set interpolation function calculates the utility value of the new participant group, and updates the lookup table according to the calculation result.

Specifically, the principle of calculating the utility value of the interpolation function is to estimate the utility value of the new participant group with the utility value of the subcombination that has been calculated. The calculation formula of the interpolation function is V(S∪{j+1})=interpolate (v _S ,v _N ,S,N), according to the calculation formula, get the utility estimate of the new participant group, that is, get the utility estimate corresponding to the new participant group S∪{j+1}, and update the lookup table v_lut [S∪{j+1}]=V(S∪{j+1}). Among them, the implementation method of the interpolation function interpolate( ) is:

In the calculation formula of the above-mentioned interpolation function, v _S , v _N , the sub-combination S and the full combination N are used as the input of the function, and are calculated and output.

In some embodiments, the utility value of the new participant group is calculated using a preset model derivation method, and the lookup table is updated according to the calculated utility value of the new participant group, including: model parameters of the new participant group Carry out aggregation, and carry out model deduction on the model of the new participant group, aggregate the weight of each participant in the new participant group, and obtain the weight of the new participant group, and the model of the new participant group in the standard verification set Carry out model deduction on the above, calculate the real utility value of the new participant group, and use the real utility value to update the lookup table.

Specifically, the model deduction uses the formula V(S∪{j+1})=V(M _S∪{j+1} ( ^t ))=V(Agg(S∪{j+1})), according to the calculation The formula obtains the real utility value of the new participant group and updates the lookup table v_lut[S∪{j+1}]=V(S∪{j+1}).

In some embodiments, based on the updated lookup table, calculate the marginal contribution value of the participant, and judge whether the marginal contribution value of the participant is converged, and when the judgment result is yes, use the converged marginal contribution value as the participant's Contribution value, including: according to the utility value of the participant in the updated lookup table after joining the sub-combination, using the preset Shapley value calculation formula to calculate the marginal contribution value of the participant; and according to whether the marginal contribution value of the participant Convergence, when the judgment result is convergence, the marginal contribution value after convergence is used as the contribution value of the participant; where the contribution value is used to indicate the contribution of the participant to the joint learning model trained in the aggregation cycle in the joint learning.

Specifically, according to the updated v_lut lookup table, calculate the marginal contribution value of participant i

if

It means that the marginal contribution value of participant i in the current full permutation combination has converged. At this time, directly set

As the contribution value of each participant i in the t-th aggregation cycle

Otherwise, let k=k+1, and regenerate a new full permutation combination P, repeat the steps of the above embodiment, and calculate the contribution value of each participant in the full permutation combination P.

Here, for the setting of the threshold θ, θ can represent the condition for judging whether the Monte Carlo method converges, and in practical applications, θ=1e-3~1e-5 can be set _.

Based on the calculation process of the contribution value of each participant in the full permutation combination P in the above embodiment, all the full permutation combinations are cycled once to obtain the contribution value of each participant i in all T aggregation periods, and the accumulation is obtained by the participant Contribution value of i to the joint model

Furthermore, Shapley Value is a method for fairly distributing benefits based on the average marginal contribution of individual i joining combination S, and its computational complexity is O( ²ⁿ ), where n is the total number of individuals. Its calculation formula is:

Further, the first truncation threshold λ is set in the following manner: Let the marginal gain of the final joint model utility function relative to the initial model be Δ _U =|V(M ^(T) )-V(M ⁽⁰⁾ )| , where T is the total communication round, which can be set to λ=Δ _U *0.01. The second truncation threshold η can be used to represent the error level of the contribution value, and can be set as η=1e-3˜1e-5.

In some embodiments, based on the framework of joint learning, multiple participant groups are generated, and a set of participant groups composed of multiple participant groups is determined, and weights of the participant groups are calculated, including: determining all participants in the joint learning The number of participants is enumerated in order from the least to the largest to obtain multiple participant groups, and the set of multiple participant groups is used as the set of participant groups, and based on the number of participants in each participant group The quantity calculates the weight corresponding to each participant group; wherein, the weight is used to represent the probability that the participant group appears in the participant group set.

Specifically, the following will describe in detail the participants of the joint learning and the process of constructing a participant group set in combination with a specific embodiment, which may specifically include the following:

Assuming that there are

N participants

1, 2, ... i...n-1, n joint learning, the training has been aggregated for T cycles, recording each aggregation cycle t during the training process, and the local data uploaded by each participant i Model M _i ^(t) , and the joint model M ^(t) after central aggregation, initialize the model M ⁽⁰⁾ , have an evaluation function or utility function V( ) for model performance (such as accuracy, loss, etc.), and the joint Learning model aggregation method Agg(·), thresholds λ, η; where λ represents the first truncated threshold, and η represents the second truncated threshold.

Further, first enumerate all possible participant groups Ps=[(1,),(2,),(3,)...,(1 ,2),(1,3),(2,3),…P,…N]; for each subcombination S with 0, 1, 2,…n-1 participants, calculate the weight w _{|S |} = |S|! (|N|-|S| ^-1 )! /|N|! .

It should be noted that each participant group corresponds to a subcombination S mentioned above. When calculating the weight of the subcombination S, it is based on the number of participants in each subcombination. The party corresponds to the elements in a set, that is, the weight corresponding to the participant group is calculated according to the number of elements in the participant group, and the weight corresponding to each sub-combination can be considered as the probability of the sub-combination appearing in the overall participant group.

In some embodiments, determining the aggregation period in the joint learning, obtaining the utility change value corresponding to the joint learning model before and after the aggregation period and establishing a lookup table includes: for each aggregation period in the joint learning, determining the joint learning corresponding to the aggregation period The initial utility value and final utility value of the model, calculate the difference between the final utility value and the initial utility value, use the difference as the utility change value, and establish a lookup table containing all participant groups corresponding to the aggregation period; for the lookup table Perform an initialization operation so that the initial utility values of other participant groups in the lookup table except the empty set participant group and the full set participant group are 0; wherein, the lookup table is used to store the corresponding utility values of all participant groups.

Specifically, calculate the utility value corresponding to the joint learning model before and after each aggregation period, and establish a lookup table, that is, for each aggregation period, you can first calculate the final utility value of this aggregation period and this aggregation period The calculation process of the initial utility value and the final utility value of the aggregation cycle will be described in detail below in conjunction with a specific embodiment, which may specifically include the following:

For each aggregation period t, calculate _v N=V(M ^(t) ), v ₀ =V(M ^(t-1) ), and establish a lookup table v_lut={():v0,(1,): 0,(2,):0,(3,):0…,(1,2):0,(1,3):0,(2,3):0,…N:vN}, where, v _N indicates the final utility value of the joint model after the end of the current aggregation period, and v ₀ indicates the utility value of the joint model after the previous aggregation period corresponding to the current aggregation period. Of course, v ₀ can also be understood as the current aggregation period before the start of the current aggregation period. For the initial utility value of the period, the difference in different expressions does not constitute a limit to the essential meaning of v ₀ , and the above two expressions are equivalent.

Further, when performing the initialization operation on the lookup table, except for the participant groups corresponding to the empty set ( ) and the full set N, the utility values of other participant groups in the participant group set Ps are set to be 0. By establishing a v_lut lookup table and using the lookup table to cache the utility value of the participant group, so as to record the calculated utility value, the amount of calculation can be reduced for subsequent contribution value calculations and double calculations can be avoided.

In some embodiments, judging whether to calculate the contribution value of each participant in the aggregation period according to the utility change value includes: comparing the utility change value of the corresponding joint learning model before and after the aggregation period with the preset first cut-off threshold, when When the utility change value is less than the first truncation threshold, and the utility change values corresponding to multiple consecutive rounds of aggregation periods are all less than the first truncation threshold, it is judged that the contribution value of each participant in the aggregation period is 0; otherwise, the contribution value of each participant in the aggregation period The contribution value is recalculated.

Fig. 10 is a schematic structural diagram of another device for determining the contribution degree of a participant in joint learning provided by an embodiment of the present disclosure. As shown in Figure 10, the device for determining the contribution of the participants in the joint learning includes:

The generation module 1001 is configured to generate a plurality of participant groups based on a federated learning architecture, and determine a participant group set composed of a plurality of participant groups, and calculate weights of the participant groups, wherein each participant group contains at least two parties;

The establishment module 1002 is configured to determine the aggregation period in the joint learning, obtain the utility change value corresponding to the joint learning model before and after the aggregation period and establish a lookup table, and judge whether to calculate the contribution value of each participant in the aggregation period according to the utility change value;

The judging module 1003 is configured to, when the judging result is yes, use the participant groups in the participant group set to randomly generate a full permutation combination, and generate multiple subcombinations according to the order of the participants in the participant group in the full permutation combination, Calculate the estimated value of the marginal contribution value when the participant joins the sub-combination, and judge whether to use the interpolation function to determine the utility value of the new participant group formed after the participant joins the sub-combination according to the estimated value of the marginal contribution value and the weight of the participant group Calculation;

The update module 1004 is configured to use an interpolation function to calculate the utility value of the new participant group when the judgment result is yes, and calculate the utility value of the new participant group by using a preset model derivation method when the judgment result is no, and The lookup table is updated according to the calculated utility value of the new party group;

The calculation module 1005 is configured to calculate the marginal contribution value of the participant based on the updated lookup table, and judge whether the marginal contribution value of the participant is converged, and when the judgment result is yes, use the converged marginal contribution value as the participant's When the judgment result is no, a new full permutation combination is generated until the contribution values of all converged participants are calculated, and the contribution of the participants in the joint learning is determined according to the contribution values.

In some embodiments, the judging module 1003 in FIG. 10 divides the full permutation combination into multiple subcombinations according to the order of the participants, and determines the next participant corresponding to the last participant in the subcombination in the full permutation combination, and calculates the following An estimate of a party's contribution margin when it joins a subgroup.

In some embodiments, the judging module 1003 in FIG. 10 calculates the product of the estimated value of the marginal contribution value and the weight of the participant group, and compares the product with the preset second truncation threshold; when the products are all less than or equal to the second When the threshold is truncated, it is judged to use the interpolation function to calculate the utility value of the new participant group; otherwise, the utility value of the new participant group is calculated using the preset model derivation method.

In some embodiments, the update module 1004 in FIG. 10 uses a preset interpolation function to update the new participant's The utility value of the group is calculated, and the lookup table is updated according to the calculation result.

In some embodiments, the update module 1004 in FIG. 10 aggregates the model parameters of the new participant group, performs model deduction on the model of the new participant group, and aggregates the weight of each participant in the new participant group , get the weight of the new participant group, model the model of the new participant group on the standard verification set, calculate the real utility value of the new participant group, and use the real utility value to update the lookup table.

In some embodiments, the generating module 1001 in FIG. 10 determines all the participants in the joint learning, enumerates the participants in order from the least to the largest to obtain multiple participant groups, and forms the multiple participant groups The set of is used as a set of participant groups, and the weight corresponding to each participant group is calculated based on the number of participants in each participant group; wherein, the weight is used to represent the probability of a participant group appearing in the participant group set.

In some embodiments, the first judgment module 1002 in FIG. 10 determines the initial utility value and the final utility value of the joint learning model corresponding to the aggregation period for each aggregation period in the joint learning, and calculates the difference between the final utility value and the initial utility value. The difference between them is used as the utility change value, and a lookup table containing all participant groups corresponding to the aggregation period is established; the lookup table is initialized so that the empty set participant group and the full set participant group are excluded from the lookup table The initial utility value of other participant groups other than the group is 0; wherein, the lookup table is used to store the utility values corresponding to all participant groups.

In some embodiments, the building module 1002 in FIG. 10 compares the utility change value of the corresponding joint learning model before and after the aggregation period with the preset first cut-off threshold. When the utility change value is less than the first cut-off threshold, and the When the utility change values corresponding to the aggregation period are all less than the first cut-off threshold, it is judged that the contribution value of each participant in the aggregation period is 0; otherwise, the contribution value of each participant in the aggregation period is recalculated.

In some embodiments, the calculation module 1005 of FIG. 10 calculates the marginal contribution value of the participant according to the utility value of the participant in the updated lookup table after adding the sub-combination, using the preset Shapley value calculation formula; and according to Whether the marginal contribution value of the participant is converged. When the judgment result is convergent, the converged marginal contribution value is used as the contribution value of the participant; where the contribution value is used to represent the joint learning of the participant for the aggregation cycle training in the joint learning Contribution of the model.

Fig. 11 is a schematic flowchart of a joint learning training method provided by an embodiment of the present disclosure. The federated learning training method in FIG. 11 may be executed by a federated learning server. As shown in Figure 11, the joint learning training method may specifically include:

S1101. In the current round of aggregation cycle of the joint learning, obtain the local model obtained by the initial model training performed by the participants of the joint learning according to the local data, and perform an aggregation operation on the local models of the participants to obtain the joint model;

S1102. Using the preset joint learning contribution value algorithm, calculate the contribution value of each participant to the joint model in the current aggregation cycle, and obtain the joint learning contribution value corresponding to each participant;

S1103. Obtain the initial index of each participant, perform a fusion operation on the joint learning contribution value and the initial index, and obtain the contribution index of each participant, where the initial index is used to represent the initial contribution of the participant to the joint learning;

S1104. Calculate the training rounds of the participants in the next aggregation cycle according to the contribution index, so that the participants can train the local model based on the training rounds in the next aggregation cycle until the training of the joint model reaches the preset Target.

Specifically, each participant corresponds to a node in the joint learning framework, and each node corresponds to a participant. Participants can be sensors, rotating machinery equipment, Internet of Things devices, IOT devices, PCs, tablet computers, smart phones, Smart wearable devices, etc., can also be objects such as companies or factories. Each participant has a client terminal of a joint learning participant, but the participant is not limited to the above-mentioned devices or clients. The federated learning framework also has a node that provides services for the client (that is, the server). The server can be a server for performing aggregation operations. The server can coordinate multiple clients to perform joint learning to obtain a joint learning model. The server may be an independent physical server, or a server cluster or cloud computing server composed of multiple physical servers.

Further, it should be noted that the process of calculating the contribution value and calculating the contribution index according to the contribution value in the joint learning of the present disclosure, and determining the training round of the participant in the next round of aggregation cycle according to the contribution index can be performed in The operation to be performed after the aggregation period corresponding to each round in the joint learning process is completed; The degree guides the training rounds of the participants in the next aggregation cycle, so that the participants with higher contributions can participate more in the training and aggregation of the joint model, thereby speeding up the convergence speed and model performance of the joint model.

According to the technical solution provided by the embodiments of the present disclosure, during the current round of aggregation period of the federated learning, the local model obtained by the initial model training performed by the participants of the federated learning according to the local data is obtained, and the aggregation operation is performed on the local models of the participants to obtain Joint model; use the preset joint learning contribution value algorithm to calculate the contribution value of each participant to the joint model in the current round of aggregation cycle, and obtain the corresponding joint learning contribution value of each participant; obtain each participant's The initial index performs the fusion operation on the joint learning contribution value and the initial index to obtain the contribution index of each participant, where the initial index is used to represent the initial contribution of the participant to the joint learning; according to the contribution index, the next The training rounds in one round of aggregation cycle, so that the participants will train the local model based on the training rounds in the next round of aggregation cycle until the training of the joint model reaches the preset goal. The disclosure can automatically adjust the training rounds of the participants in the next aggregation cycle according to the performance of the participants in the joint learning, thereby improving the convergence speed and model performance of the joint learning.

In some embodiments, during the current aggregation period of the joint learning, the local models obtained by the initial model training performed by the participants of the joint learning according to the local data are obtained, including: at the beginning of the current aggregation period of the joint learning, each participant The initial model is downloaded from the preset aggregation server, and the participants use the local data to perform an initial round of local model training on the initial model; where the initial round is the number of rounds obtained by initializing the parameters of the current round of aggregation cycle.

Specifically, the current round of aggregation period here can also be understood as the aggregation period of completed training, that is, the last round of aggregation period. In the previous round of aggregation cycle, each participant in the federated learning downloaded the initial model from the aggregation server, and performed several rounds of local model training operations on the initial model using local data, that is, each participant downloaded the initial model from the aggregation server, The initial model is trained for several rounds with local data to obtain a local model corresponding to each participant.

Further, the initial round may be the initialization model training round preset by the server for each participant. For example, the local training rounds of each participant may be recorded as local_epoch_1, local_epoch_2, local_epoch_i, ... local_epoch_n, where n is an integer greater than or equal to 0. In practical applications, the initial rounds of each participant can be set to the same value.

Further, each participant can upload their local model weight M_i (or gradient update) to the aggregation server after completing the local model training of the last round of aggregation cycle, so that the aggregation server can use the federated learning model aggregation algorithm to The local model of each party is aggregated to obtain the joint model M_global, and the parameters of the joint model are sent to each participant.

In some embodiments, after applying the joint learning model aggregation algorithm to aggregate the local models of all participants, apply the participant contribution value algorithm based on the Shapley value to calculate the contribution of each participant to the joint model in the current round of aggregation cycle. The contribution value (that is, the Shapley value). The calculation process of the contribution value corresponding to each participant will be described in detail below with reference to the accompanying drawings and specific embodiments. FIG. 12 is a schematic flowchart of the calculation of the joint learning contribution value of the participants provided by an embodiment of the present disclosure. As shown in Figure 12, the calculation process of the joint learning contribution value of the participant may specifically include:

S1201. Construct all participant combinations according to the participants in the joint learning, and calculate the weight corresponding to each participant combination;

S1202. Obtain the utility change value corresponding to the joint model before and after the current round of aggregation period, and judge whether to calculate the joint learning contribution value of each participant in the current round of aggregation period according to the utility change value;

S1203, when the judgment result is yes, select any combination of participants, and calculate the marginal contribution value corresponding to each participant in the combination of participants;

S1204, according to the marginal contribution value and the weight, judge the calculation method of the utility value of the participant combination, so as to choose to calculate the utility value of the participant combination by using an interpolation function or model deduction;

S1205. Update the predetermined lookup table according to the utility value of the combination of participants, and calculate the joint learning contribution value of each participant to the joint model based on the updated lookup table.

Specifically, firstly, according to all the participants in the joint learning, all participants are constructed in a manner from the fewest to the largest number of participants to form all participant combinations, and each participant combination corresponds to a participant group. For example, assuming that there are n participants in a joint learning, enumerate all possible participant combinations Ps=[(1,),(2,),(3,)...,( 1,2),(1,3),(2,3),…P,…N], so a subcombination S with 0, 1, 2,…n-1 participants can be obtained, that is, several Participant group. For each subcombination S with 0, 1, 2, ... n-1 participants, calculate the weight w _|S| = |S|! (|N|-|S|-1)! /|N|! .

Here, each participant combination corresponds to a sub-combination S. When calculating the weight of the sub-combination S, it is based on the number of participants in each sub-combination. In the participant combination, one participant corresponds to a set elements, that is, calculate the weight corresponding to the participant combination according to the number of elements in the participant combination, and the weight corresponding to each sub-combination can be considered as the probability of the sub-combination appearing in the overall participant combination.

Further, calculate the utility value corresponding to the joint model at the beginning and after the aggregation period, and establish a lookup table, that is, for each aggregation period, you can first calculate the final utility value of this aggregation period and the initial utility of this aggregation period value. For example, for the aggregation period t, calculate v _N =V(M ^(t) ), v ₀ =V(M ^(t-1) ), and establish a lookup table, where v _N represents the joint model after the current aggregation period v ₀ represents the utility value of the joint model after the previous aggregation period corresponding to the current aggregation period. Of course, v ₀ can also be understood as the initial utility value of the current aggregation period before the start of the current aggregation period.

Further, when it is judged that it is necessary to calculate the contribution value of each participant in the current round of aggregation period, a participant combination P is sequentially selected from all the participant combinations Ps in the current round. For each sub-participant j in P, P can be divided into two subsets {j} and S=P\{j}, that is, P=S∪{j}; calculate the marginal contribution generated by adding j to S, When calculating the marginal contribution, the actual marginal contribution should be Δ _{j_real} =v _S∪{j} -v _S =V(S∪{j})-V(S)=V(P)-V(S); However, the value of V(P) cannot be determined for the time being, so the scaling principle is used to enlarge V(P) to v _N , so that the marginal contribution of j is estimated as Δ _{j_est} =v _N -v _S =v_lut[N]- v_lut[S], and v _S is added to the list VS_hist (corresponding to another cache table), since v _S has been calculated in the previous combination P', here only needs to be obtained from the lookup table v_lut, no need to calculate V( M _S ^(t) ).

In some embodiments, according to the marginal contribution value and the weight, the calculation method of the utility value of the participant combination is judged, so as to choose to use the interpolation function or the way of model deduction to calculate the utility value of the participant combination, including: according to the participant's The product between the marginal contribution value and the weight of the participant combination, compare the product with the preset cut-off threshold, and when the product corresponding to each participant in the participant combination is less than or equal to the cut-off threshold, choose to use the interpolation function Otherwise, choose to use the model deduction method to calculate the utility value of the participant combination.

Specifically, it may be determined whether to calculate the utility value of the participant combination according to the relationship between the product of the participant's marginal contribution value and the weight of the participant corresponding to the participant combination and the cut-off threshold. In practical applications, calculate the product |Δ _{j_est} *w _|S| _{j_est} *w _|S| |≤η*|vN-v0|, it is judged to use the utility value of the sub-combination to estimate the utility value of the participant combination P, otherwise, it is judged to use the preset model derivation method to calculate the participant combination P utility value.

Further, when using the interpolation function to calculate the utility value of the participant combination, based on the utility value of the participant combination calculated in the historical iteration process and the corresponding utility value when the participant combination is a complete set of participant combinations, the preset interpolation value The function estimates the utility value of the participant combination, obtains the estimated value corresponding to the utility value of the participant combination, and updates the lookup table according to the estimated value.

Further, when using model deduction to calculate the utility value of the participant combination, the lookup table is updated according to the calculated utility value of the participant combination, the model parameters corresponding to the participant combination are aggregated, and the participant combination Carry out model deduction for the corresponding model, aggregate the weight of each participant in the participant combination to obtain the weight of the participant combination, perform model deduction on the standard verification set for the participant combination, and calculate the real utility value of the participant combination , to update the lookup table with the real utility value.

In some embodiments, before obtaining the initial index of each participant and performing the fusion operation on the joint learning contribution value and the initial index, the method further includes: for each participant in the joint learning, according to each participant's report The quality of local data, the amount of local data and/or the cost of joint learning, calculate the initial index corresponding to each participant, and normalize the initial index. The contribution of the joint model training of .

Specifically, when determining the initial index corresponding to the participant, it can be calculated according to the data quality uploaded by each participant or the joint learning cost and other parameters. For example: for a joint learning model training task, suppose there are P1, P2, Pi, ...Pn has a total of N participants, and each participant has a preset or calculated index Qi (the index value after normalization processing), which satisfies the following conditions: if Qi is larger, the participant i pairs of joint The greater the "value" or "function" of the model.

Further, the initial index can be determined according to the data quality or data volume reported by the participants, or by the negative or reciprocal of the joint learning cost reported by the participants, if the data quality, data quantity, distribution and other attributes reported by the participants Consistent, then the lower the participant's quotation for the joint learning cost, the more beneficial it is for the training task. Therefore, the initial indicators of each participant can be used as the predicted value of the contribution to the joint model training in the current aggregation period.

In some embodiments, the fusion operation is performed on the joint learning contribution value and the initial index to obtain the contribution index of each participant, including: obtaining the initial index list generated by the initial index corresponding to all the participants of the joint learning, and The contribution value list generated by the joint learning contribution value corresponding to all the participants; based on the initial index list and the contribution value list, the fusion operation is performed on the joint learning contribution value and the initial index of each participant to obtain the fused contribution index , generate a list of contribution indicators according to the contribution indicators corresponding to all participants.

Specifically, the initial index list [Q1, Q2, Qi..., Qn] is generated according to the initial index corresponding to all participants, and the contribution value list [φ1, φ2 , φi, ... φn]. According to the above-mentioned initial index list and contribution value list obtained, based on the data in the two lists, the contribution index G corresponding to each participant is calculated by weighted average or averaging. The contribution index G generates a new contribution index list [G1, G2, Gi..., Gn], which is a contribution list obtained by fusing real contribution values. Therefore, the contribution index G can represent the degree of contribution of the participants to the joint model training in the last aggregation cycle. The larger the value of the contribution index G, the higher the contribution of the participants.

In some embodiments, calculating the training rounds of the participants in the next aggregation period according to the contribution index includes: generating a mapping function according to the pre-established mapping relationship between the contribution index and the training round, and calculating The contribution index corresponding to the participant is used as the input of the mapping function, and the mapping function is used to calculate the training round corresponding to the participant, and the training round is used as the aggregation round when the participant trains the local model in the next round of aggregation cycle.

Specifically, before calculating the training round of the next round of aggregation period corresponding to each participant, the contribution index G corresponding to the participant calculated according to the foregoing embodiment can be calculated and the contribution index G can be combined with the local training round A positively correlated mapping relationship (for example, a linear or nonlinear mapping relationship) is established for local_epoch_i times, and a mapping function local_epoch_i'=f(Gi) is generated according to the established mapping relationship.

Furthermore, when calculating the training round of a participant based on the mapping function, the contribution index Gi of the participant i in the last aggregation cycle can be used as the input of the mapping function, and the local_epoch_i of the participant i can be automatically calculated by using the mapping function ' (that is, the training round of the participant in the next round of aggregation cycle). In practical applications, the larger the contribution index Gi corresponding to the participant, the greater the calculated value of local_epoch_i', which can reach the upper limit of the set local training round; the smaller the contribution index Gi, the larger the value of local_epoch_i' The smaller it is, the lowest can be reduced to 0, which means that the participant will not participate in joint model training or aggregation in the next round of aggregation cycle.

Finally, each participant downloads the joint model M from the aggregation server, and applies a new training round local_epoch_i', and uses local data to train the local model of the local_epoch_i' round, and repeats the operations in the above embodiments until the joint model reaches convergence or reaches The set maximum aggregation rounds, and finally a joint model that has been trained for T aggregation cycles.

According to the technical solutions provided by the embodiments of the present disclosure, after each round of model aggregation, the embodiments of the present disclosure calculate the aggregation round of the next round of aggregation cycle according to the contributions made by the participants in this round to the joint model, so that Instruct all participants to apply new aggregation rounds for local model training, so that participants with high contributions can participate more in the training and aggregation of the joint model, thereby speeding up the convergence speed and model performance of the joint model.

Fig. 13 is a schematic structural diagram of a joint learning and training device provided by an embodiment of the present disclosure. As shown in Figure 13, the joint learning training device includes:

The aggregation module 1301 is configured to obtain the local model obtained by the initial model training performed by the participants of the joint learning according to the local data in the current round of aggregation period of the joint learning, and perform an aggregation operation on the local models of the participants to obtain the joint model;

The calculation module 1302 is configured to use the preset joint learning contribution value algorithm to calculate the contribution value of each participant to the joint model in the current round of aggregation period, and obtain the corresponding joint learning contribution value of each participant;

The fusion module 1303 is configured to obtain the initial index of each participant, and perform a fusion operation on the joint learning contribution value and the initial index to obtain the contribution index of each participant, wherein the initial index is used to represent the participant's contribution to the joint learning initial contribution;

The training module 1304 is configured to calculate the training rounds of the participants in the next aggregation period according to the contribution index, so that the participants can train the local model based on the training rounds in the next aggregation period until the joint model The training achieves the preset target.

In some embodiments, when the aggregation module 1301 of FIG. 13 starts the current round of aggregation cycle of joint learning, each participant downloads the initial model from the preset aggregation server, and the participant uses local data to perform the initial round of initial model aggregation. Local model training; where, the initial round is the number of rounds obtained by initializing the parameters of the current round of aggregation cycle.

In some embodiments, the calculation module 1302 in FIG. 13 constructs all participant combinations according to the participants in the joint learning, and calculates the weight corresponding to each participant combination; obtains the utility change corresponding to the joint model before and after the current round of aggregation period Value, according to the utility change value, judge whether to calculate the joint learning contribution value of each participant in the current aggregation cycle; when the judgment result is yes, select any combination of participants, and calculate the The corresponding marginal contribution value; according to the marginal contribution value and weight, judge the calculation method of the utility value of the participant combination, so as to choose to use interpolation function or model deduction to calculate the utility value of the participant combination; according to the utility value of the participant combination The predetermined lookup table is updated, and based on the updated lookup table, the joint learning contribution value of each participant to the joint model is calculated.

In some embodiments, the calculation module 1302 in FIG. 13 compares the product with a preset cut-off threshold according to the product of the marginal contribution value of the participant and the weight of the participant combination. When each participant in the participant combination When the corresponding products are all less than or equal to the cut-off threshold, choose to use the interpolation function to calculate the utility value of the participant combination, otherwise, choose to use the model deduction method to calculate the utility value of the participant combination.

In some embodiments, before the fusion module 1303 in FIG. 13 acquires the initial index of each participant and performs the fusion operation on the joint learning contribution value and the initial index, for each participant in the joint learning, according to each participant Reported local data quality, local data volume and/or joint learning cost, calculate the initial index corresponding to each participant, and normalize the initial index. Contribution to periodic joint model training.

In some embodiments, the fusion module 1303 in FIG. 13 obtains the initial indicator list generated by the initial indicators corresponding to all participants in the joint learning, and the contribution value list generated by the joint learning contribution values corresponding to all the participants; based on The list of initial indicators and the list of contribution values, the fusion operation is performed on the joint learning contribution value of each participant and the initial indicators to obtain the fused contribution indicators, and the contribution indicator list is generated according to the contribution indicators corresponding to all participants.

In some embodiments, the training module 1304 in FIG. 13 generates a mapping function according to the pre-established mapping relationship between the contribution index and the training round, and uses the calculated contribution index corresponding to the participant as the input of the mapping function, Use the mapping function to calculate the training rounds corresponding to the participants, and use the training rounds as the aggregation rounds when the participants train the local model in the next round of aggregation cycle.

FIG. 14 is a schematic structural diagram of an electronic device 14 provided by an embodiment of the present disclosure. As shown in FIG. 5 , the electronic device 14 of this embodiment includes: a processor 1401 , a memory 1402 , and a computer program 1403 stored in the memory 1402 and capable of running on the processor 1401 . When the processor 1401 executes the computer program 1403, the steps in the foregoing method embodiments are implemented. Alternatively, when the processor 1401 executes the computer program 1403, the functions of the modules/units in the foregoing device embodiments are implemented.

Exemplarily, the computer program 1403 can be divided into one or more modules/units, and one or more modules/units are stored in the memory 1402 and executed by the processor 1401 to complete the present disclosure. One or more modules/units may be a series of computer program instruction segments capable of accomplishing specific functions, and the instruction segments are used to describe the execution process of the computer program 1403 in the electronic device 14 .

The electronic equipment 14 may be electronic equipment such as desktop computers, notebooks, palmtop computers, and cloud servers. The electronic device 14 may include but not limited to a processor 1401 and a memory 1402 . Those skilled in the art can understand that FIG. 5 is only an example of the electronic device 14, and does not constitute a limitation to the electronic device 14. It may include more or less components than shown in the figure, or combine certain components, or different components. , for example, an electronic device may also include an input and output device, a network access device, a bus, and the like.

The processor 1401 may be a central processing unit (Central Processing Unit, CPU), or other general-purpose processors, digital signal processors (Digital Signal Processor, DSP), application specific integrated circuits (Application Specific Integrated Circuit, ASIC), on-site Field-Programmable Gate Array (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor, or the processor may be any conventional processor, or the like.

The storage 1402 may be an internal storage unit of the electronic device 14 , for example, a hard disk or a memory of the electronic device 14 . The memory 1402 can also be an external storage device of the electronic device 14, for example, a plug-in hard disk equipped on the electronic device 14, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD) card, a flash memory card ( Flash Card), etc. Further, the memory 1402 may also include both an internal storage unit of the electronic device 14 and an external storage device. The memory 1402 is used to store computer programs and other programs and data required by the electronic device. The memory 1402 can also be used to temporarily store data that has been output or will be output.

Those skilled in the art can clearly understand that for the convenience and brevity of description, only the division of the above-mentioned functional units and modules is used for illustration. In practical applications, the above-mentioned functions can be assigned to different functional units, Completion of modules means that the internal structure of the device is divided into different functional units or modules to complete all or part of the functions described above. Each functional unit and module in the embodiment may be integrated into one processing unit, or each unit may exist separately physically, or two or more units may be integrated into one unit, and the above-mentioned integrated units may adopt hardware It can also be implemented in the form of software functional units. In addition, the specific names of the functional units and modules are only for the convenience of distinguishing each other, and are not used to limit the protection scope of the present application. For the specific working process of the units and modules in the above-mentioned system, reference may be made to the corresponding process in the foregoing method embodiments, and details are not repeated here.

In the above-mentioned embodiments, the descriptions of each embodiment have their own emphases, and for parts that are not detailed or recorded in a certain embodiment, refer to the relevant descriptions of other embodiments.

Those skilled in the art can appreciate that the units and algorithm steps of the examples described in conjunction with the embodiments disclosed herein can be implemented by electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are executed by hardware or software depends on the specific application and design constraints of the technical solution. Skilled artisans may implement the described functionality using different methods for each particular application, but such implementation should not be considered beyond the scope of the present disclosure.

In the embodiments provided in the present disclosure, it should be understood that the disclosed apparatus/computer equipment and methods may be implemented in other ways. For example, the device/computer device embodiments described above are only illustrative, for example, the division of modules or units is only a logical function division, and there may be other division methods in actual implementation, and multiple units or components can be Incorporation may either be integrated into another system, or some features may be omitted, or not implemented. In another point, the mutual coupling or direct coupling or communication connection shown or discussed may be through some interfaces, and the indirect coupling or communication connection of devices or units may be in electrical, mechanical or other forms.

A unit described as a separate component may or may not be physically separated, and a component displayed as a unit may or may not be a physical unit, that is, it may be located in one place, or may be distributed to multiple network units. Part or all of the units can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

In addition, each functional unit in each embodiment of the present disclosure may be integrated into one processing unit, each unit may exist separately physically, or two or more units may be integrated into one unit. The above-mentioned integrated units can be implemented in the form of hardware or in the form of software functional units.

If an integrated module/unit is realized in the form of a software function unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the present disclosure realizes all or part of the processes in the methods of the above embodiments, and can also be completed by instructing related hardware through computer programs. The computer programs can be stored in computer-readable storage media, and the computer programs can be processed. When executed by the controller, the steps in the above-mentioned method embodiments can be realized. A computer program may include computer program code, which may be in source code form, object code form, executable file, or some intermediate form or the like. The computer-readable medium may include: any entity or device capable of carrying computer program code, recording medium, U disk, removable hard disk, magnetic disk, optical disk, computer memory, read-only memory (Read-Only Memory, ROM), random access Memory (Random Access Memory, RAM), electrical carrier signal, telecommunication signal and software distribution medium, etc. It should be noted that the content contained in computer readable media may be appropriately increased or decreased according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, computer readable media may not Including electrical carrier signals and telecommunication signals.

The above embodiments are only used to illustrate the technical solutions of the present disclosure, rather than to limit them; although the present disclosure has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: it can still be described in the foregoing embodiments Modifications to the technical solutions recorded, or equivalent replacements for some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present disclosure, and should be included in this disclosure. within the scope of protection.

Claims

A method for determining the contribution of a participant in joint learning, characterized by comprising:

Determine the participants of the joint learning, construct all participant combinations based on the participants, and calculate the weight corresponding to each participant combination;

Determine the utility change value corresponding to the joint learning model before and after the aggregation period and establish a lookup table, and judge whether to calculate the contribution value of the participant in the aggregation period according to the utility change value;

When the judgment result is yes, select a participant combination from all the participant combinations according to a predetermined order, and estimate the marginal contribution value of each participant in the selected participant combination, according to the estimation result and The weight of the combination of participants, judging whether to use an interpolation function to calculate the utility value of the combination of participants;

When the judgment result is yes, use the interpolation function to calculate the utility value of the combination of participants; when the judgment result is no, use the preset model deduction method to calculate the utility value of the combination of participants, and calculate updating the lookup table with the utility value of the combination of the participants;

Select each participant combination in turn until the utility values of all the participant combinations are calculated, and update the lookup table using the utility values of all the participant combinations to obtain a final updated lookup table, so that based on the The final updated lookup table is used to calculate the contribution value of each participant, and the contribution degree of the participant in the joint learning is determined according to the contribution value.
The method according to claim 1, wherein the constructing all participant combinations based on the participants and calculating the weight corresponding to each participant combination includes: according to all the participants in the joint learning According to the number of participants from the fewest to the most, the combination of all participants is constructed. The combination of all participants includes a plurality of combinations of participants. According to the number of participants in each combination of participants Calculating the weight; wherein, the weight is used to represent the probability that the participant combination appears in all the participant combinations;

Alternatively, calculating the utility value of the participant combination by using a preset model derivation method, and updating the lookup table according to the calculated utility value of the participant combination includes: The corresponding model parameters are aggregated, and the model corresponding to the participant combination is modeled, and the weight of each participant in the participant combination is aggregated to obtain the weight of the participant combination. The combination of the participants performs model deduction on the standard verification set, calculates the real utility value of the combination of the participants, and uses the real utility value to update the look-up table;

Alternatively, the calculating the contribution value of each participant based on the final updated lookup table includes: according to the utility value of each participant combination in the final updated lookup table, using The preset Shapley value calculation formula calculates the contribution value corresponding to each of the participants, wherein the contribution value is used to represent the joint learning of the participants in the joint learning for the aggregation cycle training Contribution of the model.
The method according to claim 1, wherein determining the utility change value corresponding to the joint learning model before and after the aggregation period and establishing a lookup table includes: for each of the aggregation periods, determining the initial utility of the aggregation period value and the final utility value, using the difference between the final utility value and the initial utility value as the utility change value, and establishing a lookup table corresponding to the aggregation period that contains all the combinations of the participants; The lookup table performs an initialization operation, so that the initial utility value of other participant combinations in the lookup table except the empty set participant combination and the full set participant combination is 0; wherein, the lookup table is used to store all The utility value corresponding to the combination of the participants;

Alternatively, the judging whether to calculate the contribution value of the participant in the aggregation period according to the utility change value includes: comparing the utility change value of the aggregation period with a preset first cut-off threshold, when the When the utility change value of the aggregation cycle is less than the first cut-off threshold, and the utility change values of the aggregation cycle for multiple consecutive rounds are all smaller than the first cut-off threshold, the contribution value of each participant in the aggregation cycle is judged is 0; otherwise, calculate the contribution value of each participant in the aggregation period.
The method according to claim 1, characterized in that, selecting a participant combination from all the participant combinations according to a predetermined order, and making a marginal contribution to each participant in the selected participant combination Values are estimated, including:

According to the arrangement order of the participant combinations in all the participant combinations, the participant combinations are selected in turn, and the participants in the selected participant combinations are iterated in turn, so as to The marginal contribution value generated when the party joins the participant combination is estimated to obtain the estimated value of the marginal contribution value.
The method according to claim 4, wherein, according to the estimation result and the weight of the participant combination, judging whether to use an interpolation function to calculate the utility value of the participant combination includes: calculating the participation The product of the estimated value of the marginal contribution value corresponding to the party and the weight of the participant corresponding to the participant combination, and comparing the product with the preset second cut-off threshold; when each of the participant combination When the products corresponding to the participants are all less than or equal to the preset second cut-off threshold, it is judged that an interpolation function is used to calculate the utility value of the participant combination; otherwise, the preset model derivation method is used to calculate the The utility value of the combination of the above participants;

The calculation of the utility value of the participant combination by using an interpolation function includes: based on the utility value of the participant combination calculated in the historical iteration process, and the corresponding utility value when the participant combination is a complete set of participant combinations, Estimate the utility value of the participant combination by using a preset interpolation function to obtain an estimated value corresponding to the utility value of the participant combination, and update the lookup table according to the estimated value.
A device for determining the contribution degree of a participant in joint learning, characterized in that it includes:

The construction module is configured to determine the participants of the joint learning, construct all participant combinations based on the participants, and calculate the weight corresponding to each participant combination;

The first judging module is configured to determine the utility change value corresponding to the joint learning model before and after the aggregation period and establish a lookup table, and judge whether to calculate the contribution value of the participant in the aggregation period according to the utility change value;

The second judging module is configured to, when the judging result is yes, select a participant combination from all the participant combinations in a predetermined order, and make a marginal contribution to each participant in the selected participant combination value, and judge whether to use an interpolation function to calculate the utility value of the participant combination according to the estimation result and the weight of the participant combination;

The update module is configured to use an interpolation function to calculate the utility value of the participant combination when the judgment result is yes, and to calculate the utility value of the participant combination by using a preset model derivation method when the judgment result is no, and updating the lookup table according to the calculated utility value of the participant combination;

The calculation module is configured to sequentially select each combination of participants until the utility values of all the combinations of participants are calculated, and update the lookup table by using the utility values of all combinations of participants to obtain the final updated A lookup table, so as to calculate the contribution value of each participant based on the finally updated lookup table, and determine the contribution of the participant in the joint learning according to the contribution value.
A method for determining the contribution of a participant in joint learning, characterized by comprising:

determining all participant combinations based on the participants in the joint learning, and calculating the weight of each participant combination in the all participant combinations;

determining a first utility value of the joint model before the start of the current aggregation period, and a second utility value of the joint model after the end of the current aggregation period, calculating a utility change value based on the first utility value and the second utility value, and Establish a lookup table; wherein, the utility change value is used to determine whether to calculate the contribution value of each participant in the current aggregation period;

When judging and calculating the contribution value, select a participant combination from all the participant combinations, and calculate the marginal contribution value corresponding to each participant in the participant combination, according to the marginal contribution value and the Weight, judging whether to use the first estimation method or the second estimation method to estimate the utility value of the participant combination;

determining the estimation result of the utility value of the participant combination by the first estimation method or the second estimation method, using the estimation result to update the lookup table, and iteratively estimating in turn to obtain each participant combination The utility value of , and obtain the final lookup table obtained after updating according to the utility value, so as to use the final lookup table to calculate the contribution value of the participant.
The method according to claim 7, wherein said determining the first utility value of the joint model before the start of the current aggregation period and the second utility value of the joint model after the end of the current aggregation period is based on the first The utility value and the second utility value calculate the utility change value, and establish a lookup table, including: determining the first utility value and the second utility value corresponding to the joint model before and after the current aggregation period begins, and calculating The difference between the second utility value and the first utility value, using the difference as the utility change value, and establishing a lookup table corresponding to the current aggregation period that contains all the combinations of the participants ; Perform an initialization operation on the lookup table, so that the initial utility value of other participant combinations in the lookup table except the empty set participant combination and the full set participant combination is 0; wherein, the lookup table is used for storing utility values corresponding to all combinations of participants;

Alternatively, the utility change value is used to judge whether to calculate the contribution value of each participant in the current aggregation period, including: comparing the utility change value of the joint model corresponding to the current aggregation period with a preset first cut-off threshold , when the utility change value is less than the first cut-off threshold, it is judged that the contribution value of each participant in the current aggregation period is 0; otherwise, the contribution value of each participant in the current aggregation period is re- calculate.
The method according to claim 7, wherein the selecting a participant combination from all the participant combinations, and calculating the corresponding marginal contribution value of each participant in the participant combination includes: The order of arrangement of each participant combination in the all participant combinations is to select a participant combination from the all participant combinations in turn, and randomly select a participant from the participant combination; according to the participant A sub-combination is divided from the combination of participants, the marginal contribution value generated when the participant joins the sub-combination is calculated, and the participants in the combination of participants are iterated sequentially, so as to calculate and obtain each The marginal contribution value corresponding to the participant; wherein, the sub-combination is a set of participants other than the randomly selected participant in the participant combination;

According to the marginal contribution value and the weight, judging to use the first estimation method or the second estimation method to estimate the utility value of the participant combination includes: calculating the marginal contribution value of the participant and the The product between the weights of the participant combination to which the participant belongs, and compare the product with the preset second cut-off threshold; when the product corresponding to each participant in the participant combination is When it is less than or equal to the second cut-off threshold, it is judged to use the first estimation method to estimate the utility value of the participant combination; otherwise, use the second estimation method to estimate the utility value of the participant combination Make an estimate.
The method according to claim 9, characterized in that estimating the utility value of the participant combination by using the first estimation method comprises: obtaining the utility value of the participant combination from the lookup table The utility value corresponding to the sub-combination, calculate the average value, maximum value or minimum value of the utility value corresponding to the sub-combination, and use the calculated average value, maximum value or minimum value as the utility value of the participant combination an estimated value, and updating the lookup table according to the estimated value;

Alternatively, the estimating the utility value of the participant combination by using the second estimation method includes: aggregating model parameters corresponding to the participant combination, and modeling the model corresponding to the participant combination Deduction, the weight of each of the participants in the combination of participants is aggregated to obtain the weight of the combination of participants, model deduction is performed on the combination of participants on a standard verification set, and the calculation of the combination of the participants is obtained combined real utility value, using the real utility value to update the lookup table.
A device for determining the contribution degree of a participant in joint learning, characterized in that it includes:

A determination module configured to determine all participant combinations based on the participants in the joint learning, and calculate the weight of each participant combination in the all participant combinations;

A building module configured to determine a first utility value of the joint model before the start of the current aggregation period, and a second utility value of the joint model after the end of the current aggregation period, based on the first utility value and the second utility value Calculating the utility change value, and establishing a lookup table; wherein, the utility change value is used to judge whether to calculate the contribution value of each participant in the current aggregation period;

The judging module is configured to, when judging and calculating the contribution value, select a participant combination from all the participant combinations, and calculate the marginal contribution value corresponding to each participant in the participant combination, according to the The marginal contribution value and the weight determine to use the first estimation method or the second estimation method to estimate the utility value of the participant combination;

The calculation module is configured to determine the estimation result of the utility value of the participant combination by the first estimation method or the second estimation method, use the estimation result to update the look-up table, and iteratively estimate in turn to obtain each A utility value of the participant combination, and a final lookup table obtained after updating according to the utility value is obtained, so as to use the final lookup table to calculate the contribution value of the participant.
A method for determining the contribution of a participant in joint learning, characterized by comprising:

Based on the framework of joint learning, multiple participant groups are generated, and a set of participant groups composed of multiple participant groups is determined, and weights of the participant groups are calculated, wherein each of the participant groups includes at least two participants;

Determine the aggregation period in the joint learning, obtain the utility change value corresponding to the joint learning model before and after the aggregation period and establish a lookup table, and judge whether to calculate the contribution value of each participant in the aggregation period according to the utility change value;

When the judgment result is yes, use the participant groups in the participant group set to randomly generate a full permutation combination, and generate multiple subcombinations according to the order of the participants in the participant group in the full permutation combination, calculate The estimated value of the marginal contribution value when the participant joins the sub-combination, according to the estimated value of the marginal contribution value and the weight of the participant group, it is judged whether to use an interpolation function to add the participant to the sub-combination Calculate the utility value of the new participant group formed;

When the judgment result is yes, use the interpolation function to calculate the utility value of the new participant group, when the judgment result is no, use the preset model derivation method to calculate the utility value of the new participant group, and calculate updating the lookup table with the utility value of the new party group;

Based on the updated lookup table, calculate the marginal contribution value of the participant, and judge whether the marginal contribution value of the participant is converged, and when the judgment result is yes, use the converged marginal contribution value as the participant's Contribution value, when the judgment result is no, generate a new full permutation combination until the calculation of the contribution value of all the participants after convergence, and determine the contribution of the participant in the joint learning according to the contribution value contribution.
The method according to claim 12, characterized in that, generating a plurality of sub-combinations according to the order of the participants in the participant group in the full arrangement combination, and calculating the marginal contribution value of the participants when they join the sub-combinations Estimates, including:

Divide the full permutation combination into multiple subcombinations according to the order of the participants, and determine the next participant corresponding to the last participant in the subcombination in the full permutation combination, and calculate the next participant Estimates of the contribution margins when a party joins the subportfolio.
The method according to claim 12, characterized in that, based on the updated lookup table, the contribution margin of the participant is calculated, and it is judged whether the contribution contribution of the participant converges, and when the judgment result is yes When , take the converged marginal contribution value as the contribution value of the participant, including:

According to the utility value of the participant in the updated lookup table after joining the sub-combination, the preset Shapley value calculation formula is used to calculate the marginal contribution value of the participant; and according to the participant Whether the marginal contribution value of the converged value is converged, and when the judgment result is converged, the converged marginal contribution value is used as the contribution value of the participant;

Wherein, the contribution value is used to represent the contribution of the participant to the joint learning model trained in the aggregation period in the joint learning.
A device for determining the contribution degree of a participant in joint learning, characterized in that it includes:

The generation module is configured to generate a plurality of participant groups based on a joint learning architecture, and determine a participant group set composed of a plurality of the participant groups, and calculate the weights of the participant groups, wherein each of the participant groups The party group contains at least two parties;

The establishment module is configured to determine the aggregation period in the joint learning, obtain the utility change value corresponding to the joint learning model before and after the aggregation period, and establish a lookup table, and judge whether to calculate each in the aggregation period according to the utility change value. Contribution value of participants;

The judging module is configured to randomly generate a full permutation combination by using the participant groups in the participant group set when the judgment result is yes, and according to the order of the participants in the participant group in the full permutation combination Generate a plurality of sub-combinations, calculate the estimated value of the marginal contribution value when the participant joins the sub-combination, and judge whether to use an interpolation function to evaluate the participants according to the estimated value of the marginal contribution value and the weight of the participant group The utility value of the new participant group formed after adding the sub-combination is calculated;

The update module is configured to use an interpolation function to calculate the utility value of the new participant group when the judgment result is yes, and to calculate the utility value of the new participant group by using a preset model derivation method when the judgment result is no value, and update the lookup table according to the calculated utility value of the new participant group;

The calculation module is configured to calculate the marginal contribution value of the participant based on the updated lookup table, and judge whether the marginal contribution value of the participant is converged, and when the judgment result is yes, the converged marginal contribution value As the contribution value of the participant, when the judgment result is no, a new full permutation combination is generated until the contribution value of all converged participants is calculated, and the participant is determined according to the contribution value The degree of contribution in the joint learning.
A joint learning training method, characterized in that, comprising:

In the current round of aggregation cycle of the joint learning, the local model obtained by the initial model training of the participants of the joint learning according to the local data is obtained, and the aggregation operation is performed on the local models of the participants to obtain the joint model;

Using a preset joint learning contribution value algorithm, calculate the contribution value of each of the participants to the joint model in the current round of aggregation period, and obtain the joint learning contribution value corresponding to each of the participants;

Obtain the initial index of each of the participants, perform a fusion operation on the joint learning contribution value and the initial index, and obtain the contribution index of each of the participants, wherein the initial index is used to represent the contribution of the participant to The initial contribution of joint learning;

calculating the training rounds of the participant in the next aggregation period according to the contribution index, so that the participant can train the local model based on the training rounds in the next aggregation period, Until the training of the joint model reaches the preset target.
The method according to claim 16, characterized in that, in the current round of aggregation period of the joint learning, obtaining the local model obtained by the participants of the joint learning from the initial model training based on the local data comprises: When the current round of aggregation period starts, each of the participants downloads the initial model from the preset aggregation server, and the participant uses local data to perform an initial round of local model training on the initial model; wherein, the initial The round is the round value obtained by initializing the parameters of the current round of aggregation period;

Alternatively, by using the preset joint learning contribution value algorithm, the contribution value of each of the participants to the joint model in the current round of aggregation period is calculated to obtain the joint learning corresponding to each of the participants. Contribution value, including: according to the participants in the joint learning, construct all the participant combinations, and calculate the weight corresponding to each of the participant combinations; obtain the utility change value corresponding to the joint model before and after the current round of aggregation period , judging whether to calculate the joint learning contribution value of each of the participants in the current aggregation cycle according to the utility change value; when the judgment result is yes, select any combination of the participants, and calculate the The marginal contribution value corresponding to each of the participants in the participant combination; according to the marginal contribution value and the weight, the calculation method of the utility value of the participant combination is judged, so as to choose to use an interpolation function or model Calculating the utility value of the combination of participants by means of deduction; updating the predetermined lookup table according to the utility value of the combination of participants, and based on the updated lookup table, calculating the contribution of each participant to the The joint learning contribution value of the joint model.

Or, according to the marginal contribution value and the weight, the calculation method of the utility value of the participant combination is judged, so as to choose to calculate the utility value of the participant combination by using an interpolation function or model derivation, Including: according to the product of the marginal contribution value of the participant and the weight of the participant combination, comparing the product with a preset cut-off threshold, when each participant in the participant combination When the corresponding products are all less than or equal to the cut-off threshold, choose to use an interpolation function to calculate the utility value of the participant combination, otherwise, choose to use model deduction to calculate the utility value of the participant combination;

Alternatively, before acquiring the initial index of each of the participants and performing the fusion operation on the joint learning contribution value and the initial index, the method further includes: for each of the participants in the joint learning , according to the local data quality, local data volume and/or joint learning cost reported by each of the participants, calculate the initial index corresponding to each of the participants, and perform normalization processing on the initial index, the The initial indicator is used to represent the estimated contribution of the participant to the joint model training of the current aggregation cycle;

Alternatively, the calculating the training round of the participant in the next round of aggregation period according to the contribution index includes: generating a mapping function according to the pre-established mapping relationship between the contribution index and the training round, and The calculated contribution index corresponding to the participant is used as the input of the mapping function, and the training round corresponding to the participant is calculated by using the mapping function, and the training round is used as the participant's next round Aggregation rounds when training the local model within an aggregation cycle.
A joint learning training device, characterized in that it comprises:

The aggregation module is configured to obtain the local model obtained by the participants of the joint learning according to the local data for the initial model training in the current round of aggregation cycle of the joint learning, and perform an aggregation operation on the local models of the participants to obtain the joint model;

The calculation module is configured to use a preset joint learning contribution value algorithm to calculate the contribution value of each of the participants to the joint model in the current round of aggregation period, and obtain the corresponding joint learning contribution value;

The fusion module is configured to obtain an initial index of each of the participants, perform a fusion operation on the joint learning contribution value and the initial index, and obtain a contribution index of each of the participants, wherein the initial index is used for Characterize the initial contribution of the participants to the joint learning;

A training module configured to calculate the training rounds of the participant in the next round of aggregation period according to the contribution index, so that the participant will, in the next round of aggregation period, based on the training rounds The local model is trained until the training of the joint model reaches a preset target.
An electronic device, comprising a memory, a processor, and a computer program stored in the memory and operable on the processor, the processor implements the method according to claim 1 when executing the program.
A computer-readable storage medium storing a computer program, wherein the computer program implements the method according to claim 1 when executed by a processor.