WO2023071106A1

WO2023071106A1 - Federated learning management method and apparatus, and computer device and storage medium

Info

Publication number: WO2023071106A1
Application number: PCT/CN2022/089694
Authority: WO
Inventors: 李泽远; 王健宗
Original assignee: 平安科技（深圳）有限公司
Priority date: 2021-10-26
Filing date: 2022-04-27
Publication date: 2023-05-04
Also published as: CN113947215A

Abstract

The present application relates to the technical field of artificial intelligence. Provided are a federated learning management method and apparatus, and a computer device and a storage medium. The method comprises: a plurality of participating terminals training respective local databases by means of a preset federated model, so as to obtain a model parameter corresponding to each participating terminal; training a preset joint model by means of each model parameter, and recording data of contributions of each model parameter to the joint model; on the basis of the contribution data and a preset credibility score consensus mechanism model, performing credit scoring on each participating terminal; and according to a credit score of each participating terminal, performing reward and punishment management on each participating terminal.

Description

Federal learning management method, device, computer equipment and storage medium

This application claims the priority of the Chinese patent application with application number 202111249348.0 and titled "Federal Learning Management Method, Device, Computer Equipment, and Storage Medium" filed with the China Patent Office on October 26, 2021, the entire contents of which are incorporated by reference incorporated in this application.

technical field

The embodiment of the present application relates to the field of federated learning, especially a federated learning management method, device, computer equipment and storage medium.

Background technique

For the risk control of health insurance, the insurance company calculates whether the credit value of the policyholder meets the insurance requirements through the history of major diseases of the policyholder, medical history, and the health status statement submitted by the policyholder. In cases where the sex cannot be judged, the insurance company cannot publish the user's data to obtain verification from a third-party organization, or directly aggregate the data of the medical institution into a model to determine the true credit value of the policyholder.

The inventor of the present application realized in the research that the traditional blockchain consensus mechanism based on federated learning adopts the Byzantine Fault Tolerance consensus algorithm (Practical Byzantine Fault Tolerance, PBFT). Participants' behavior evaluation, and the inability to quantify and calculate the contribution of participants after the training is over.

application content

The embodiment of the present application provides a federated learning management method, device, computer equipment, and storage medium that can evaluate the behavior of participants according to the consensus process, and quantify and calculate the contribution of participants after training.

In order to solve the above technical problems, a technical solution adopted by the embodiment created by this application is to provide a federated learning management method,

Including: multiple participating terminals train their respective local databases through the preset federated model, and obtain a model parameter corresponding to each participating terminal; train the preset joint model through each model parameter, and record each model Participate in the contribution data of the joint model; based on the contribution data and the preset reputation scoring consensus mechanism model, perform reputation scoring on each participating terminal; Carry out reward and punishment management.

In order to solve the above technical problems, the embodiment of the present application also provides a federated learning management device, including: a training module, used for multiple participating terminals to train their respective local databases through a preset federated model, and obtain the corresponding a model parameter; the training module is also used to train the preset joint model through each model parameter, and record the contribution data of each model participating in the joint model; the scoring module is used to The contribution data and the preset reputation scoring consensus mechanism model are used to perform reputation scoring on each participating terminal; the management module is used to perform reward and punishment management on each participating terminal according to the reputation scoring of each participating terminal.

In order to solve the above-mentioned technical problems, an embodiment of the present application also provides a computer device, which includes a memory and a processor, where computer-readable instructions are stored in the memory, and when the computer-readable instructions are executed by the processor, the The processor executes the federated learning management method:

Multiple participating terminals train their respective local databases through the preset federated model, and obtain a model parameter corresponding to each participating terminal;

Train the preset joint model through each model parameter, and record the contribution data of each model participant to the joint model;

Perform reputation scoring on each participating terminal based on the contribution data and the preset reputation scoring consensus mechanism model;

According to the reputation score of each participating terminal, reward and punishment management is performed on each participating terminal.

In order to solve the above-mentioned technical problems, an embodiment of the present application further provides a storage medium storing computer-readable instructions, wherein, when the computer-readable instructions are executed by one or more processors, one or more processors execute the The federated learning management approach described:

Description of drawings

The above and/or additional aspects and advantages of the present application will become apparent and easy to understand from the following description of the embodiments in conjunction with the accompanying drawings, wherein:

Fig. 1 is one of the schematic flow charts of the federated learning management method of a specific embodiment of the present application;

Fig. 2 is the second schematic flow diagram of the federated learning management method of a specific embodiment of the present application;

FIG. 3 is the third schematic flow diagram of a federated learning management method in a specific embodiment of the present application;

FIG. 4 is a schematic diagram of a consensus mechanism model of reputation scoring in a specific embodiment of the present application;

FIG. 5 is the fourth schematic flow diagram of a federated learning management method in a specific embodiment of the present application;

FIG. 6 is the fifth schematic flow diagram of a federated learning management method in a specific embodiment of the present application;

FIG. 7 is the sixth schematic flow diagram of a federated learning management method in a specific embodiment of the present application;

FIG. 8 is the seventh schematic flow diagram of a federated learning management method in a specific embodiment of the present application;

FIG. 9 is a schematic diagram of the basic structure of a federated learning management device according to an embodiment of the present application;

FIG. 10 is a block diagram of a basic structure of a computer device according to an embodiment of the present application.

Detailed ways

Embodiments of the present application are described in detail below, examples of which are shown in the drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below by referring to the figures are exemplary only for explaining the present application, and are not construed as limiting the present application.

Those skilled in the art will understand that unless otherwise stated, the singular forms "a", "an", "said" and "the" used herein may also include plural forms. It should be further understood that the word "comprising" used in the specification of the present application refers to the presence of the features, integers, steps, operations, elements and/or components, but does not exclude the presence or addition of one or more other features, Integers, steps, operations, elements, components, and/or groups thereof. In addition, "and/or" in the specification and claims means at least one of the connected objects, and the character "/" generally means that the related objects are an "or" relationship.

Those skilled in the art can understand that, unless otherwise defined, all terms (including technical terms and scientific terms) used herein have the same meanings as commonly understood by those of ordinary skill in the art to which this application belongs. It should also be understood that terms, such as those defined in commonly used dictionaries, should be understood to have meanings consistent with their meaning in the context of the prior art, and unless specifically defined as herein, are not intended to be idealized or overly Formal meaning to explain.

The embodiments of the present application may acquire and process relevant data based on artificial intelligence technology. Among them, artificial intelligence (AI) is a theory, method, technology and application system that uses digital computers or machines controlled by digital computers to simulate, extend and expand human intelligence, perceive the environment, acquire knowledge and use knowledge to obtain the best results. .

Artificial intelligence basic technologies generally include technologies such as sensors, dedicated artificial intelligence chips, cloud computing, distributed storage, big data processing technology, operation/interaction systems, and mechatronics. Artificial intelligence software technology mainly includes computer vision technology, robotics technology, biometrics technology, speech processing technology, natural language processing technology, and machine learning/deep learning.

In related technologies, various insurance systems determine whether the credit value of the policyholder meets the insurance requirements by judging the relevant information of the policyholder. Under normal circumstances, the data of the policyholder can be trained based on the federated learning model to expand the data dimension of the policyholder to determine whether the credit value of the policyholder meets the insurance requirements.

To determine whether the credit value of the policyholder meets the insurance requirements in related technologies, it is mainly implemented through the following methods:

In the traditional health insurance risk control scenario, each insurance system calculates whether the credit value of the insured meets the insurance requirements through the history of the insured’s history of major disease visits and the health status statement submitted by the insured, but there is a lack of dimensionality of the insured’s data , the authenticity of the data cannot be judged, etc., each insurance system (insurance company) cannot publish its own user data to obtain verification from a third-party organization, or directly aggregate the data of multiple medical institutions together for modeling.

Usually, by introducing a federated learning model, each insurance system can train the relevant information of the policyholder in the local database to realize the expansion of the data dimension of the policyholder. Each participant in the federated learning model needs to rely on the central node to update or issue parameters. If the central node fails or acts maliciously, the results of the entire federated learning collaborative training will be affected.

Since the original data cannot be transmitted, we can only transmit some intermediate data of the model, such as gradient information. But in fact, even if the gradient information is leaked, there is still a risk that the original data will be deduced. And there is no guarantee that every participant in federated learning is honest. Because each participant may have different motivations. Then when we talk about dishonest participants, there are two types: one is malicious, and the other is innocent but curious. The so-called malicious participant is that he may come to poison the model, such as deliberately transmitting some wrong data to harm the interests of other participants, while the curious participant will not harm the interests of other participants, but he All the interaction data he collects will be analyzed to try to deduce the original data of other parties.

Of course, federated learning in related technologies also has some other issues such as data transmission efficiency. Model training involves a large number of calculations, and the joint modeling of all parties will involve a large amount of data interaction. For example, in gradient descent, each gradient iteration involves communication costs. Therefore, communication efficiency is also a challenge that federated learning will encounter during the implementation process. In addition, there are issues such as uneven sample distribution from institution to institution, and so on.

If it is possible to upload the parameters after model training of each insurance system to the blockchain, through peer-to-peer communication, we can get rid of the dependence on the central server. At the same time, the consensus mechanism of the blockchain can recognize the contribution of each participant to reward or punish, and the source of malicious behavior can be traced afterwards, so that federated learning can reduce malicious nodes (training with invalid, virus data) or selfish nodes ( Do not actively provide data resources, only ask for resources from other participants) to participate in model training.

In view of the problems existing in the above-mentioned implementation methods, this application thinks of a federated learning management method, which can judge whether it is qualified for federated learning training according to the credibility of each participating terminal before training; participating terminals can supervise each other during training. The honesty of participating terminals during the training process is voted and scored, and their own credibility is maintained to prevent information tampering; after training, participants are rewarded and punished based on their credibility scores.

In addition, the behavior evaluation of participating terminals is carried out according to the consensus process, and the contribution to the participants is quantified and calculated after the training, which improves the accuracy of each participating terminal in obtaining and judging user data.

As shown in Figure 1, it is a schematic flow diagram of a federated learning management method provided in this embodiment, including S201 to S204:

S201. Multiple participating terminals train their respective local databases through a preset federated model to obtain a model parameter corresponding to each participating terminal.

Exemplarily, the above-mentioned multiple participating terminals may be multiple institutions (or companies) such as health big data institutions, medical institutions, and insurance institutions, and each participating terminal has its own local database, which includes its own user information .

Exemplarily, each participating terminal trains user information through a preset federated model based on its own local database, so as to obtain a model parameter corresponding to each participating terminal.

It can be understood that the above-mentioned preset federated model is an existing technology. After introducing federated learning, each participating terminal conducts training in the local database to expand the data dimension of the policyholder. In federated learning, each participant needs to rely on The central node updates or issues parameters.

In a possible implementation manner, a model parameter corresponding to each participating terminal may also be obtained through the following steps.

Exemplarily, as shown in FIG. 2, the above step S201 may include the following steps S201a and S201b:

S201a. Each participating terminal among the plurality of participating terminals respectively uses the preset federated model to train a local database to obtain model parameters and weight values corresponding to each participating terminal.

Exemplarily, each participating terminal performs data model training on the local database through the preset federated model, establishes the model, first numerically processes the information corresponding to each user in the participating terminal database, and then screens out relevant information from it. High feature information, so as to obtain the model parameters and weight values corresponding to each participating terminal.

It should be noted that federated learning can enable all participants to cooperate to complete the training of a data model. The trained model is based on the data of all participants, but the participants will not disclose their own data. Raw data.

Exemplarily, the above preset federated models may be: horizontal federated learning, vertical federated learning, federated transfer learning, and so on.

S201b. Each of the multiple participating terminals uploads corresponding model parameters and weight values to the blockchain.

Exemplarily, after each participating terminal uses the preset federated model to train the local database and obtains the model parameters and weight values corresponding to each participating terminal, it can upload the model parameters and weight values to the same zone Blockchain (shared database).

For example, health big data institutions, medical institutions, and insurance institutions can use their respective databases to train locally with a preset federated model, obtain initial model parameters and weight values, and upload them to the blockchain.

As an example, take the field of auto insurance as an example. Because each driver has different driving habits, some drivers may use their cars a lot, while some drivers may park their cars in underground garages all year round; some drivers may drive It is better to get used to it. Some drivers may like to speed and engage in some dangerous driving behaviors. It is also insured for one year, so can we design different insurance premiums for different users? For example, users who use the car for a long time have higher insurance premiums than users who use the car for less time, and users with bad driving habits are more expensive than users with good driving habits. High premiums. In this way, insurance companies can also reduce risks, increase premiums for users with a higher probability of accidents, and also block some bad users. By performing federated learning training on user information in each participating terminal, the corresponding model parameters and weight values are obtained.

In specific implementation, each participating terminal needs to create corresponding task configuration information based on the federation model. Specifically, each participating terminal can determine and create the task configuration information of the federated model task by responding to the user's federated learning setting operation; wherein, the task configuration information of the federated model task includes but is not limited to: task type, engine framework , automatic parameter tuning algorithm, early termination algorithm, feature engineering information and methods, and data preprocessing methods and other information.

After each participating terminal determines the task configuration information corresponding to the federated model task, each participating terminal sends the task configuration information to the blockchain, so that the blockchain can obtain the task configuration information of multiple participating terminals participating in the federated model. Since the task configuration information does not involve data security and privacy issues, each participating terminal can send the task configuration information to the blockchain without encryption.

S202. Train the preset joint model through each model parameter, and record the contribution data of each model participating in the joint model.

Exemplarily, after the block chain receives the model parameters uploaded by each participating terminal, these model parameters are integrated, and then joint training is performed to obtain joint model parameters (global model) corresponding to multiple participating terminals, and determine each The contribution data of participating terminals.

Exemplarily, in the process of training the preset joint model through various model parameters in the blockchain, the user behavior of each participating terminal during the training process, the contribution to the joint model and the consensus voting results can be recorded and other information.

Exemplarily, as shown in FIG. 3, the above step S202 may include the following steps S202a to S202d:

S202a. Concatenate the model parameters of the multiple participating terminals to generate federated parameters.

Exemplarily, the blockchain can integrate model parameters of multiple participating terminals to obtain a concatenated model parameter, thereby generating federated parameters.

It should be noted that the above-mentioned federation parameters are model parameters (ie, global model parameters) jointly obtained from model parameters of multiple participating terminals.

S202b. Initialize parameters of the joint model according to the federation parameters, and train the initialized joint model according to preset training samples to generate feature vectors.

Exemplarily, the blockchain initializes the model training configuration information (that is, the parameters of the joint model) in the federated learning task according to the federated parameters, and executes the model training operation of the federated learning task based on the initialized model training configuration information, Generate the corresponding eigenvectors.

Exemplarily, in the initialized model training configuration information, feature engineering information for model training operations is determined, and according to the feature engineering information, user data samples are subjected to feature processing to obtain model training data samples and corresponding feature vectors are generated.

S202c. Calculate the feature difference of the joint model based on the feature vector and a preset label vector.

S202d. Calculate the deviation value of each model parameter according to the characteristic difference value, and generate the contribution data according to the deviation value.

Exemplarily, the blockchain performs difference calculation according to the generated feature vector and the preset label vector to obtain the feature difference corresponding to the joint model.

Further, the blockchain recalculates the deviation value corresponding to each model parameter according to the obtained characteristic difference value and the model parameter corresponding to each participating terminal, so as to generate the corresponding contribution data of each participating terminal according to the corresponding deviation value of each participating terminal.

S203. Perform reputation scoring on each participating terminal based on the contribution data and the preset reputation scoring consensus mechanism model.

Exemplarily, the Practical Byzantine Fault Tolerance of Credibility Evaluation (CE-PBFT) evaluates user behavior according to the consensus process, sets the reputation score for voting dynamic weight adjustment, and performs joint training tasks with the parameters after joint model training. Generate a global model.

As shown in Figure 4, it is a model diagram of the reputation scoring consensus mechanism corresponding to the federated learning management method provided by the embodiment of this application. The reputation scoring consensus mechanism performs reputation evaluation on each participating terminal according to the determined contribution data corresponding to each participating terminal Scoring, each participating terminal corresponds to a reputation score.

Exemplarily, the contribution data includes: user behavior of each participating terminal, contribution to the joint model, and consensus voting results. As shown in FIG. 5, the above step S203 may include the following steps S203a and S203b:

S203a. Input the user behavior of each participating terminal, the contribution to the joint model, and the consensus voting result into the reputation scoring consensus mechanism model.

Exemplarily, the block chain records the user behavior of each participating terminal during the training process, the contribution to the joint model and the consensus voting results (that is, the user behavior in the training process of the historical records, the contribution to the joint model and consensus voting results, if there is no historical record, it will be calculated from this voting). After uploading the parameters this time, each participating terminal begins to vote. If a participating terminal gives up voting, the reputation score will be reduced. If it is lower than the scoring threshold, it will not be eligible for federated model learning and training.

Exemplarily, the blockchain inputs three pieces of information, namely user behavior of each participating terminal, contribution to the joint model, and consensus voting results, into the reputation scoring consensus mechanism model, so that the reputation scoring consensus mechanism model is The three pieces of information of each participating terminal are analyzed and processed to determine the reputation score corresponding to each participating terminal.

Exemplarily, the reputation scoring consensus mechanism calculates the corresponding reputation of each participating terminal based on the three information of each participating terminal's user behavior, contribution to the joint model, and consensus voting results, and according to the proportion of each information Score.

S203b. Read the reputation scores of the participating terminals output by the reputation scoring consensus mechanism model.

Exemplarily, after the credibility score consensus mechanism model calculates the reputation score corresponding to each participating terminal, it sends the reputation score corresponding to each participating terminal to each participating terminal.

Exemplarily, the calculation method of the credibility score consensus mechanism model for each participating terminal is shown in formula 1:

Among them, α, β, λ are parameters, and T is the score of previous historical votes.

Indicates the updated value of the participant's reputation score after voting, i and j represent different participants, and t represents the current number of votes. Others means to increase the score. If the participating terminal actively participates in voting or performs well during the training process, the score will be added, otherwise the score will be reduced.

Exemplarily, before the above step S203, as shown in FIG. 6, the federated learning management method provided by the embodiment of the present application may also include the following steps S301 and S302:

S301. Acquire global parameters of the global model.

Wherein, the global model is a model form when the joint model is trained to a converged state.

S302. Distribute the global parameter to each participating terminal, so that the federated model of each participating terminal generates a global parameter.

Exemplarily, the reputation score consensus mechanism model evaluates user behavior according to the consensus process, sets the reputation score for voting dynamic weight adjustment, and performs joint training tasks with parameters after model training to generate a global model and obtain global parameters of the global model.

Exemplarily, the block chain updates the model parameters of the global model and sends them to each participating terminal, so that the federated model of each participating terminal can obtain the global parameters.

S204. Perform reward and punishment management on each participating terminal according to the reputation score of each participating terminal.

Exemplarily, the reputation score consensus mechanism model rewards and punishes the reputation scores of each participating terminal according to the reputation scores of each participating terminal. Participating terminals with low reputation scores may have malicious behaviors and low contribution to this training. Unable to participate in the next round of federated model learning and training.

Exemplarily, as shown in FIG. 7, the above step S204 may include the following steps S204a and S204b:

S204a. Compare the credit score of each participating terminal with a preset score threshold.

S204b. When the reputation score of any participating terminal is less than the scoring threshold, prohibit the participating terminal from participating in the next round of joint training.

Exemplarily, after the reputation score consensus mechanism model of each participating terminal is obtained, the reputation score of each participating terminal is compared with a preset scoring threshold (for example, 50 points), and the reputation score of any participating terminal is less than When scoring the threshold, it can be determined that the contribution of the participating terminal is low, or there is malicious behavior, etc., and the participating terminal is prohibited from participating in the next round of joint training.

Specifically, when the reputation score of any participating terminal is between 50 and 100, it means that the participating terminal has excellent performance. Through active participation in model training, there is no malicious behavior in the process, and finally the reputation score reaches 100, and the points will be reset. If the value is 50, the scoring of the next cycle will start again. When the participating terminal has malicious behavior or passively participates in model training, the reputation score will continue to decrease, and eventually it will be lower than 50 and model training cannot be performed.

Exemplarily, after the above step S204, as shown in FIG. 8 , the federated learning management method provided by the embodiment of the present application may further include the following steps S401 and S402:

S401. Read the voting results of the participating terminals in sequence.

S402. When any participating terminal abstains from voting, reduce the reputation score of the participating terminal.

Exemplarily, the reputation scoring consensus mechanism model respectively obtains the voting status of each participating terminal, and updates the reputation scoring of each participating terminal after 2) the voting ends.

Exemplarily, when the reputation scoring consensus mechanism model cannot obtain the voting result of a certain participating terminal, it is determined that the participating terminal has given up voting, and the reputation score of the participating terminal is reduced.

The federated learning management method provided in this embodiment uses a preset federated model to train the respective local databases of multiple participating terminals to obtain a model parameter corresponding to each participating terminal, so that the preset federated learning parameters can be adjusted through each model parameter. The model is trained, and the contribution data of each model participating in the joint model is recorded during the training process. Finally, the contribution data of multiple participating terminals is analyzed through the preset reputation scoring consensus mechanism model to evaluate the reputation of each participating terminal. Scoring, so as to carry out reward and punishment management for each participating terminal according to the reputation score of each participating terminal. In order to solve problems such as federated learning relying on the central server, there may be failures or malicious behaviors, the model training parameters of each participating terminal are uploaded to the blockchain, and the credibility score consensus mechanism is used according to the results of voting scores. Rewarding or punishing participating terminals can fully mobilize the enthusiasm of participating terminals, and can also reduce the existence of participating terminals that have malicious or selfish behaviors.

It should be noted that, the federated learning management method provided in the embodiment of the present application may be executed by a federated learning management device, or a control module in the federated learning management device for executing the federated learning management method. In the embodiment of the present application, the federated learning management device provided in the embodiment of the present application is described by taking the federated learning management device executing the federated learning management method as an example.

It should be noted that, in the embodiments of the present application, the federated learning management methods shown in the drawings of the above methods are all described in conjunction with a drawing in the embodiments of the present application as an example. In specific implementation, the federated learning management method shown in the drawings of the above methods can also be implemented in combination with any other drawings shown in the above embodiments that can be combined, and will not be repeated here.

Please refer to FIG. 9 for details. FIG. 9 is a schematic diagram of the basic structure of the federated learning management device in this embodiment.

As shown in FIG. 9, a federated learning management device includes: a training module 801, used for multiple participating terminals to train their local databases through a preset federated model to obtain a model parameter corresponding to each participating terminal; The training module 801 is also used to train the preset joint model through each model parameter, and record the contribution data of each model participating in the joint model; the scoring module 802 is used to train the joint model based on the contribution data and The preset reputation scoring consensus mechanism model performs reputation scoring on each participating terminal; the management module 803 is configured to perform reward and punishment management on each participating terminal according to the reputation scoring of each participating terminal.

In some manners, the training module 801 is specifically used for each of the multiple participating terminals to use the preset federated model to train the local database to obtain a model corresponding to each participating terminal parameters and weight values; the device also includes: an upload module 804; the upload module 804 is used for each participating terminal in the plurality of participating terminals to upload corresponding model parameters and weight values to the block chain.

In some manners, the training module 801 is specifically configured to concatenate the model parameters of the multiple participating terminals to generate federated parameters; The parameters of the model are initialized, and the initialized joint model is trained according to the preset training samples to generate a feature vector; the training module 801 is specifically further configured to, based on the feature vector and the preset label vector, Calculate the characteristic difference of the joint model; the training module 801 is specifically configured to calculate the deviation value of each model parameter according to the characteristic difference, and generate the contribution data according to the deviation value.

In some manners, the contribution data includes: user behavior of each participating terminal, contribution to the joint model, and consensus voting results; Behavior, contribution to the joint model and consensus voting results are input into the reputation scoring consensus mechanism model; the scoring module 802 is also specifically used to read the output of the credibility scoring consensus mechanism model The reputation score of each participating terminal.

In some manners, the device further includes: an acquisition module 805 and a sending module 806; the acquisition module 805 is configured to acquire global parameters of the global model, wherein the global model is when the joint model is trained to a convergence state model form; the sending module 806 is configured to distribute the global parameters to the participating terminals, so that the federated models of the participating terminals generate global parameters.

In some manners, the management module 803 is specifically configured to compare the reputation score of each participating terminal with a preset scoring threshold; When the score is less than the score threshold, the participating terminals are prohibited from participating in the next round of joint training.

In some manners, the management module 803 is further configured to sequentially read the voting results of the participating terminals; the management module 803 is further configured to lower the voting results of the participating terminals when any participating terminal gives up voting. reputation score.

A computer device includes a memory and a processor, wherein computer-readable instructions are stored in the memory, and when the computer-readable instructions are executed by the processor, the processor executes the federated learning management method:

In some ways, the federated learning management method also includes:

Each of the plurality of participating terminals uses the preset federated model to train a local database to obtain model parameters and weight values corresponding to each participating terminal;

Each of the plurality of participating terminals uploads corresponding model parameters and weight values to the block chain.

In some ways, the federated learning management method also includes:

splicing the model parameters of the multiple participating terminals to generate federated parameters;

Initializing parameters of the joint model according to the federation parameters, and training the initialized joint model according to preset training samples to generate feature vectors;

calculating the feature difference of the joint model based on the feature vector and a preset label vector;

Calculate the deviation value of each model parameter according to the characteristic difference value, and generate the contribution data according to the deviation value.

In some manners, the contribution data includes: user behavior of each participating terminal, contribution to the joint model, and consensus voting results; the federated learning management method further includes:

Input the user behavior of each participating terminal, the contribution to the joint model and the consensus voting result into the reputation scoring consensus mechanism model;

Reading the reputation scores of the participating terminals output by the reputation scoring consensus mechanism model.

In some ways, the federated learning management method also includes:

Acquiring global parameters of the global model, wherein the global model is the model form when the joint model is trained to a converged state;

Distributing the global parameters to the participating terminals, so that the federated models of the participating terminals generate global parameters.

In some ways, the federated learning management method also includes:

Comparing the reputation scores of each participating terminal with a preset scoring threshold;

When the reputation score of any participating terminal is less than the scoring threshold, the participating terminal is prohibited from participating in the next round of joint training.

In some ways, the federated learning management method also includes:

Read the voting results of each participating terminal in sequence;

When any participating terminal abstains from voting, the reputation score of the participating terminal is reduced.

A storage medium storing computer-readable instructions, which, when executed by one or more processors, cause one or more processors to execute the federated learning management method:

Multiple participating terminals train their respective local databases through the preset federation model, and obtain a model parameter corresponding to each participating terminal;

In some ways, the federated learning management method also includes:

Read the voting results of each participating terminal in turn;

The federated learning management device in the embodiment of the present application may be a device, or a component, an integrated circuit, or a chip in a terminal. The device may be a mobile electronic device or a non-mobile electronic device. Exemplarily, the mobile electronic device may be a mobile phone, a tablet computer, a notebook computer, a handheld computer, a vehicle electronic device, a wearable device, an ultra-mobile personal computer (ultra-mobile personal computer, UMPC), a netbook or a personal digital assistant (personal digital assistant). assistant, PDA), etc., non-mobile electronic devices can be servers, network attached storage (Network Attached Storage, NAS), personal computer (personal computer, PC), television (television, TV), teller machine or self-service machine, etc., this application Examples are not specifically limited.

The server can be an independent server, or it can provide cloud services, cloud database, cloud computing, cloud function, cloud storage, network service, cloud communication, middleware service, domain name service, security service, content delivery network (Content Delivery Network, CDN), and cloud servers for basic cloud computing services such as big data and artificial intelligence platforms.

The federated learning management device provided in the embodiment of the present application can implement various processes implemented by the federated learning management device in the method embodiments shown in FIGS. 1 to 8 . To avoid repetition, details are not repeated here.

For the beneficial effects of the various implementations in this embodiment, refer to the beneficial effects of the corresponding implementations in the foregoing method embodiments. To avoid repetition, details are not repeated here.

The federated learning management device provided in the embodiment of the present application trains the respective local databases of multiple participating terminals through a preset federated model to obtain a model parameter corresponding to each participating terminal, so that the preset The joint model is trained, and the contribution data of each model participating in the joint model is recorded during the training process. Finally, the contribution data of multiple participating terminals is analyzed through the preset reputation scoring consensus mechanism model, so as to analyze the contribution data of each participating terminal. Reputation scoring, so as to carry out reward and punishment management for each participating terminal according to the reputation scoring of each participating terminal. In order to solve problems such as federated learning relying on the central server, there may be failures or malicious behaviors, the model training parameters of each participating terminal are uploaded to the blockchain, and the credibility score consensus mechanism is used according to the results of voting scores. Rewarding or punishing participating terminals can fully mobilize the enthusiasm of participating terminals, and can also reduce the existence of participating terminals that have malicious or selfish behaviors.

In order to solve the above technical problem, an embodiment of the present application further provides a computer device. Please refer to FIG. 10 for details. FIG. 10 is a block diagram of the basic structure of the computer device in this embodiment.

As shown in FIG. 10 , a schematic diagram of the internal structure of a computer device. The computer device includes a processor, a non-volatile storage medium, a memory and a network interface connected through a system bus. Wherein, the non-volatile storage medium of the computer device stores an operating system, a database, and computer-readable instructions, and the database can store control information sequences, and when the computer-readable instructions are executed by the processor, the processor can realize a A Federated Learning Management Approach. The processor of the computer equipment is used to provide computing and control capabilities and support the operation of the entire computer equipment. Computer-readable instructions may be stored in the memory of the computer device, and when the computer-readable instructions are executed by the processor, the processor may execute a federated learning management method. The network interface of the computer device is used for connecting and communicating with the terminal. Those skilled in the art can understand that the structure shown in FIG. 8 is only a block diagram of a partial structure related to the solution of this application, and does not constitute a limitation on the computer equipment to which the solution of this application is applied. The specific computer equipment can be More or fewer components than shown in the figures may be included, or some components may be combined, or have a different arrangement of components.

In this embodiment, the processor is used to execute the specific functions of the acquisition module 801 , the construction module 802 and the adjustment module 803 in FIG. 8 , and the memory stores program codes and various data required for executing the above modules. The network interface is used for data transmission between user terminals or servers. The memory in this embodiment stores the program codes and data required to execute all sub-modules in the federated learning management device, and the server can call the program codes and data of the server to execute the functions of all sub-modules.

The computer device provided in this embodiment trains the respective local databases of multiple participating terminals through a preset federated model, so as to obtain a model parameter corresponding to each participating terminal, so that the preset joint model can be performed through each model parameter. Training, and record the contribution data of each model participating in the joint model during the training process, and finally analyze the contribution data of multiple participating terminals through the preset reputation scoring consensus mechanism model, so as to perform reputation scoring on each participating terminal, Therefore, according to the reputation score of each participating terminal, reward and punishment management is performed on each participating terminal. In order to solve problems such as federated learning relying on the central server, there may be failures or malicious behaviors, the model training parameters of each participating terminal are uploaded to the blockchain, and the credibility score consensus mechanism is used according to the results of voting scores. Rewarding or punishing participating terminals can fully mobilize the enthusiasm of participating terminals, and can also reduce the existence of participating terminals that have malicious or selfish behaviors.

The application can be used in numerous general purpose or special purpose computer system environments or configurations. Examples: personal computers, server computers, handheld or portable devices, tablet-type devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics, network PCs, minicomputers, mainframe computers, including A distributed computing environment for any of the above systems or devices, etc. This application may be described in the general context of computer-executable instructions, such as program modules, being executed by a computer. Generally, program modules include routines, programs, objects, components, data structures, etc. that perform particular tasks or implement particular abstract data types. The application may also be practiced in distributed computing environments where tasks are performed by remote processing devices that are linked through a communications network. In a distributed computing environment, program modules may be located in both local and remote computer storage media including storage devices

The present application also provides a storage medium storing computer-readable instructions. When the computer-readable instructions are executed by one or more processors, the one or more processors execute the steps of the federated learning management method in any of the above embodiments.

Those of ordinary skill in the art can understand that realizing all or part of the processes in the methods of the above embodiments can be completed by instructing related hardware through a computer program, and the computer program can be stored in a computer-readable storage medium. During execution, it may include the processes of the embodiments of the above-mentioned methods. Wherein, the aforementioned storage medium may be a nonvolatile storage medium such as a magnetic disk, an optical disk, a read-only memory (Read-Only Memory, ROM), or a random access memory (Random Access Memory, RAM).

The present application also provides a storage medium storing computer-readable instructions, the storage medium of the computer-readable instructions may be non-volatile or volatile, and the computer-readable instructions are executed by one or more processors When, one or more processors are made to execute the steps of the federated learning management method in any of the above embodiments.

Those of ordinary skill in the art can understand that realizing all or part of the processes in the methods of the above embodiments can be completed by instructing related hardware through a computer program, the computer program can be stored in a computer-readable storage medium, and the program is executed , may include the flow of the embodiments of the above-mentioned methods. Wherein, the aforementioned storage medium may be a nonvolatile storage medium such as a magnetic disk, an optical disk, a read-only memory (Read-Only Memory, ROM), or a random access memory (Random Access Memory, RAM).

Those skilled in the art can understand that the various operations, methods, and steps, measures, and schemes in the processes that have been discussed in this application can be replaced, changed, combined, or deleted. Furthermore, the various operations, methods, and other steps, measures, and schemes in the processes that have been discussed in this application may also be replaced, changed, rearranged, decomposed, combined, or deleted. Further, steps, measures, and schemes in the prior art that have operations, methods, and processes disclosed in the present application may also be alternated, changed, rearranged, decomposed, combined, or deleted.

The above descriptions are only some implementations of the present application. It should be pointed out that for those of ordinary skill in the art, some improvements and modifications can be made without departing from the principle of the application. These improvements and modifications are also It should be regarded as the protection scope of this application.

Claims

A method for federated learning management, comprising:

Multiple participating terminals train their respective local databases through the preset federated model, and obtain a model parameter corresponding to each participating terminal;

Train the preset joint model through each model parameter, and record the contribution data of each model participant to the joint model;

Perform reputation scoring on each participating terminal based on the contribution data and the preset reputation scoring consensus mechanism model;

According to the reputation score of each participating terminal, reward and punishment management is performed on each participating terminal.
The method according to claim 1, wherein the plurality of participating terminals train their respective local databases through a preset federated model to obtain a model parameter corresponding to each participating terminal, including:

Each of the plurality of participating terminals uses the preset federated model to train a local database to obtain model parameters and weight values corresponding to each participating terminal;

Each of the plurality of participating terminals uploads corresponding model parameters and weight values to the block chain.
The method according to claim 1, wherein the preset joint model is trained through each model parameter, and the contribution data of each model participating in the joint model is recorded, including:

splicing the model parameters of the multiple participating terminals to generate federated parameters;

Initializing parameters of the joint model according to the federation parameters, and training the initialized joint model according to preset training samples to generate feature vectors;

calculating the feature difference of the joint model based on the feature vector and a preset label vector;

Calculate the deviation value of each model parameter according to the characteristic difference value, and generate the contribution data according to the deviation value.
The method according to claim 1, wherein the contribution data includes: user behavior of each participating terminal, contribution to the joint model, and consensus voting results;

Based on the contribution data and the preset reputation scoring consensus mechanism model, performing reputation scoring on each participating terminal includes:

Input the user behavior of each participating terminal, the contribution to the joint model and the consensus voting result into the reputation scoring consensus mechanism model;

Reading the reputation scores of the participating terminals output by the reputation scoring consensus mechanism model.
The method according to claim 4, wherein, before performing reputation scoring on each participating terminal based on the contribution data and the preset reputation scoring consensus mechanism model, it includes:

Acquiring global parameters of the global model, wherein the global model is the model form when the joint model is trained to a converged state;

Distributing the global parameters to the participating terminals, so that the federated models of the participating terminals generate global parameters.
The method according to claim 1, wherein said performing reward and punishment management on said participating terminals according to the reputation scores of said participating terminals includes:

Comparing the reputation scores of each participating terminal with a preset scoring threshold;

When the reputation score of any participating terminal is less than the scoring threshold, the participating terminal is prohibited from participating in the next round of joint training.
The method according to claim 1, wherein, after performing reward and punishment management on each participating terminal according to the reputation score of each participating terminal, comprising:

Read the voting results of each participating terminal in turn;

When any participating terminal abstains from voting, the reputation score of the participating terminal is reduced.
A federated learning management device, including:

The training module is used for multiple participating terminals to train their respective local databases through a preset federated model to obtain a model parameter corresponding to each participating terminal;

The training module is also used to train the preset joint model through each model parameter, and record the contribution data of each model participating in the joint model;

A scoring module, configured to perform reputation scoring on each participating terminal based on the contribution data and the preset reputation scoring consensus mechanism model;

A management module, configured to perform reward and punishment management on each participating terminal according to the reputation score of each participating terminal.
A computer device, including a memory and a processor, wherein computer-readable instructions are stored in the memory, and when the computer-readable instructions are executed by the processor, the processor executes the federated learning management method :

Multiple participating terminals train their respective local databases through the preset federated model, and obtain a model parameter corresponding to each participating terminal;

Train the preset joint model through each model parameter, and record the contribution data of each model participant to the joint model;

Perform reputation scoring on each participating terminal based on the contribution data and the preset reputation scoring consensus mechanism model;

According to the reputation score of each participating terminal, reward and punishment management is performed on each participating terminal.
The computer device according to claim 9, wherein the federated learning management method further comprises:

Each of the plurality of participating terminals uses the preset federated model to train a local database to obtain model parameters and weight values corresponding to each participating terminal;

Each of the plurality of participating terminals uploads corresponding model parameters and weight values to the block chain.
The computer device according to claim 9, wherein the federated learning management method further comprises:

splicing the model parameters of the multiple participating terminals to generate federated parameters;

Initializing parameters of the joint model according to the federation parameters, and training the initialized joint model according to preset training samples to generate feature vectors;

calculating the feature difference of the joint model based on the feature vector and a preset label vector;

Calculate the deviation value of each model parameter according to the characteristic difference value, and generate the contribution data according to the deviation value.
The computer device according to claim 9, wherein the contribution data includes: user behavior of each participating terminal, contribution to the joint model, and consensus voting results; the federated learning management method further includes:

Input the user behavior of each participating terminal, the contribution to the joint model and the consensus voting result into the reputation scoring consensus mechanism model;

Reading the reputation scores of the participating terminals output by the reputation scoring consensus mechanism model.
The computer device according to claim 12, wherein the federated learning management method further comprises:

Acquiring global parameters of the global model, wherein the global model is the model form when the joint model is trained to a converged state;

Distributing the global parameters to the participating terminals, so that the federated models of the participating terminals generate global parameters.
The computer device according to claim 9, wherein the federated learning management method further comprises:

Comparing the reputation scores of each participating terminal with a preset scoring threshold;

When the reputation score of any participating terminal is less than the scoring threshold, the participating terminal is prohibited from participating in the next round of joint training.
The computer device according to claim 9, wherein the federated learning management method further comprises:

Read the voting results of each participating terminal in sequence;

When any participating terminal abstains from voting, the reputation score of the participating terminal is reduced.
A storage medium storing computer-readable instructions, wherein, when the computer-readable instructions are executed by one or more processors, one or more processors execute the federated learning management method:

Multiple participating terminals train their respective local databases through the preset federated model, and obtain a model parameter corresponding to each participating terminal;

Train the preset joint model through each model parameter, and record the contribution data of each model participant to the joint model;

Perform reputation scoring on each participating terminal based on the contribution data and the preset reputation scoring consensus mechanism model;

According to the reputation score of each participating terminal, reward and punishment management is performed on each participating terminal.
The storage medium according to claim 16, wherein the federated learning management method further comprises:

Each participating terminal in the plurality of participating terminals uses the preset federated model to train the local database respectively, and obtains model parameters and weight values corresponding to each participating terminal;

Each of the plurality of participating terminals uploads corresponding model parameters and weight values to the block chain.
The storage medium according to claim 16, wherein the federated learning management method further comprises:

splicing the model parameters of the multiple participating terminals to generate federated parameters;

Initializing parameters of the joint model according to the federation parameters, and training the initialized joint model according to preset training samples to generate feature vectors;

calculating the feature difference of the joint model based on the feature vector and a preset label vector;

Calculate the deviation value of each model parameter according to the characteristic difference value, and generate the contribution data according to the deviation value.
The storage medium according to claim 16, wherein the contribution data includes: user behavior of each participating terminal, contribution to the joint model, and consensus voting results; the federated learning management method further includes:

Input the user behavior of each participating terminal, the contribution to the joint model and the consensus voting result into the reputation scoring consensus mechanism model;

Reading the reputation scores of the participating terminals output by the reputation scoring consensus mechanism model.
The storage medium according to claim 19, wherein the federated learning management method further comprises:

Acquiring global parameters of the global model, wherein the global model is the model form when the joint model is trained to a converged state;

Distributing the global parameters to the participating terminals, so that the federated models of the participating terminals generate global parameters.