CN114386613A

CN114386613A - Model updating method based on federal learning, information sending method and equipment

Info

Publication number: CN114386613A
Application number: CN202011121251.7A
Authority: CN
Inventors: 于路; 信伦
Original assignee: China Mobile Communications Group Co Ltd; China Mobile Communications Ltd Research Institute
Current assignee: China Mobile Communications Group Co Ltd; China Mobile Communications Ltd Research Institute
Priority date: 2020-10-19
Filing date: 2020-10-19
Publication date: 2022-04-22

Abstract

The invention provides a model updating method based on federal learning, an information sending method and equipment, belonging to the technical field of big data analysis, wherein the model updating method based on federal learning comprises the following steps: obtaining first model parameters sent by at least two target terminals, and obtaining quality evaluation results of training data of the at least two target terminals, wherein the first model parameters are updated model parameters obtained by the target terminals performing model updating on a target federal learning model based on the training data; according to the quality evaluation results of the training data of the at least two target terminals, carrying out weighted summation on the first model parameters to obtain second model parameters; and updating the model of the target federal learning model according to the second model parameter. The invention can improve the model quality and the modeling efficiency.

Description

Model updating method based on federal learning, information sending method and equipment

Technical Field

The invention relates to the technical field of big data analysis, in particular to a method and equipment for updating a model, sending information and sending equipment based on federal learning.

Background

In recent years, artificial intelligence has entered a big data era dominated by deep learning, and big data is a basic element for current artificial intelligence application. However, in practical applications, we are faced with the data status quo: small scale, fragmentation, lack of data tags, data scatter, data islanding caused by user privacy protection and data security, etc. The traditional approach is to aggregate data scattered throughout a data center where modeling is performed. With the continuous expansion of the application field and the continuous improvement of various laws and regulations for protecting the privacy of users, the traditional method for modeling after data aggregation is more and more difficult to implement, so that the federal study is born, and the basic idea of the federal study is to construct a high-performance model shared by a plurality of participants under the condition of ensuring the privacy of users and the safety of data, and aims to solve the problems of privacy protection and data isolated island.

The horizontal federal learning is an important component of the federal learning, is mainly applied to a plurality of participants, has the same data structure but insufficient data quantity, and hopefully obtains a high-performance model through combined modeling under the condition that data privacy is not mutually exposed. Besides the terminals participating in modeling, the horizontal federal study also comprises a central server (coordinator), and the central server is responsible for collecting model data of all participants, integrating and organizing all the participants to perform model optimization and updating, and finally obtaining a combined modeling model. The basic flow is as follows:

1. initial parameter sending: the central server sends model initial parameters to each terminal participating in model updating in the current round;

2. training and updating each terminal model: each terminal receives the model parameters of the central server, uses the model parameters as initial values, applies self data to carry out model training, updates the model parameters and sends the updated parameters to the central server;

3. and after receiving all the updated parameters, the central server performs weighted summation on all the parameters according to a certain rule, updates the parameters of the combined model and obtains an updated global model. The updating of the model of the current round is finished;

4. and returning to 1.

The above process is repeated until a certain convergence condition is reached.

In the modeling process of federal learning, model quality is a very important issue. In practice, it is found that the phenomena of low model quality and low modeling efficiency sometimes occur.

Disclosure of Invention

In view of this, the invention provides a model updating method, an information sending method and equipment based on federal learning, which are used for solving the problems of low quality and low modeling efficiency of the current federal learning model.

In order to solve the above technical problem, in a first aspect, the present invention provides a model updating method based on federal learning, which is applied to a server, and includes:

obtaining first model parameters sent by at least two target terminals, and obtaining quality evaluation results of training data of the at least two target terminals, wherein the first model parameters are updated model parameters obtained by the target terminals performing model updating on a target federal learning model based on the training data;

according to the quality evaluation results of the training data of the at least two target terminals, carrying out weighted summation on the first model parameters to obtain second model parameters;

and updating the model of the target federal learning model according to the second model parameter.

Optionally, the quality evaluation result is obtained by evaluating at least one of data volume, data integrity and data distribution balance of the training data.

Optionally, the target federal learning model is a home care rehabilitation management model for stroke patients.

Optionally, before the obtaining the first model parameters sent by the at least two target terminals, the method further includes:

obtaining quality evaluation results of training data of at least two terminals;

and selecting the target terminal from the terminals according to the quality evaluation results of the training data of the at least two terminals.

Optionally, the selecting the target terminal from the terminals according to the quality evaluation results of the training data of the at least two terminals includes:

selecting a first terminal from the terminals according to the quality evaluation results of the training data of the at least two terminals;

sending initial parameters of the target federated learning model to the first terminal;

receiving a first model parameter sent by the first terminal within a preset time period, wherein the first model parameter is an updated model parameter obtained by the first terminal performing model update on the target federated learning model after receiving the initial parameter;

and taking the first terminal corresponding to the first model parameter received in the preset time as the target terminal.

In a second aspect, the present invention further provides an information sending method, applied to a terminal, including:

sending a quality evaluation result of a first model parameter and training data to a server, so that the server performs weighted summation on the first model parameters sent by at least two terminals according to the quality evaluation result to obtain a second model parameter, and performing model updating on a target federated learning model according to the second model parameter;

the first model parameter is an updated model parameter obtained by the terminal performing model update on the target federated learning model based on the training data.

Optionally, the sending the quality evaluation result of the first model parameter and the training data to the server includes:

sending a quality evaluation result of the training data to the server;

receiving initial parameters of the target federal learning model sent by the server after receiving the quality evaluation result of the training data;

updating the model of the target federal learning model according to the initial parameters and the training data to obtain first model parameters;

and sending the first model parameters to the server.

In a third aspect, the present invention further provides a server, including:

the system comprises an acquisition module, a quality evaluation module and a processing module, wherein the acquisition module is used for acquiring first model parameters sent by at least two target terminals and acquiring quality evaluation results of training data of the at least two target terminals, and the first model parameters are updated model parameters obtained by the target terminals performing model updating on a target federal learning model based on the training data;

the weighted summation module is used for carrying out weighted summation on the first model parameter according to the quality evaluation results of the training data of the at least two target terminals to obtain a second model parameter;

and the global model updating module is used for updating the model of the target federal learning model according to the second model parameters.

Optionally, the server further includes:

the quality evaluation result acquisition module is used for acquiring the quality evaluation results of the training data of at least two terminals;

and the screening module is used for selecting the target terminal from the terminals according to the quality evaluation results of the training data of the at least two terminals.

Optionally, the screening module includes:

the screening unit is used for selecting a first terminal from the terminals according to the quality evaluation results of the training data of the at least two terminals;

an initial parameter sending unit, configured to send an initial parameter of the target federated learning model to the first terminal;

a first model parameter receiving unit, configured to receive a first model parameter sent by the first terminal within a preset time period, where the first model parameter is an updated model parameter obtained by the first terminal performing model update on the target federated learning model after receiving the initial parameter;

and the target terminal determining unit is used for taking the first terminal corresponding to the first model parameter received in the preset time as the target terminal.

In a fourth aspect, the present invention further provides a terminal, including:

the sending module is used for sending a quality evaluation result of the first model parameters and the training data to a server, so that the server performs weighted summation on the first model parameters sent by at least two terminals according to the quality evaluation result to obtain second model parameters, and performs model updating on a target federated learning model according to the second model parameters;

Optionally, the sending module includes:

a quality evaluation result sending unit, configured to send a quality evaluation result of the training data to the server;

the initial parameter receiving unit is used for receiving initial parameters of the target federated learning model sent by the server after receiving the quality evaluation result of the training data;

the model updating unit is used for updating the model of the target federal learning model according to the initial parameters and the training data to obtain first model parameters;

a first model parameter sending unit, configured to send the first model parameter to the server.

In a fifth aspect, the present invention further provides a server, including: a transceiver and a processor;

the transceiver is used for obtaining first model parameters sent by at least two target terminals and obtaining quality evaluation results of training data of the at least two target terminals, wherein the first model parameters are updated model parameters obtained by the target terminals performing model updating on a target federal learning model based on the training data;

the processor is used for carrying out weighted summation on the first model parameter according to the quality evaluation results of the training data of the at least two target terminals to obtain a second model parameter;

Optionally, the transceiver is configured to obtain quality evaluation results of training data of at least two terminals;

the processor is configured to select the target terminal from the terminals according to the quality evaluation results of the training data of the at least two terminals.

Optionally, the processor is configured to select a first terminal from the terminals according to a quality evaluation result of the training data of the at least two terminals;

the transceiver is used for sending the initial parameters of the target federal learning model to the first terminal;

the transceiver is further configured to receive a first model parameter sent by the first terminal within a preset time period, where the first model parameter is an updated model parameter obtained by the first terminal performing model update on the target federal learning model after receiving the initial parameter;

the processor is further configured to use the first terminal corresponding to the first model parameter received within the preset time as the target terminal.

In a sixth aspect, the present invention further provides a terminal, including: a transceiver and a processor;

the transceiver is used for sending a quality evaluation result of a first model parameter and training data to a server, so that the server performs weighted summation on the first model parameters sent by at least two terminals according to the quality evaluation result to obtain a second model parameter, and performs model updating on a target federated learning model according to the second model parameter;

Optionally, the transceiver is configured to send a quality evaluation result of the training data to the server;

the transceiver is further configured to receive initial parameters of the target federated learning model sent by the server after receiving a quality evaluation result of the training data;

the processor is used for carrying out model updating on the target federal learning model according to the initial parameters and the training data to obtain first model parameters;

the transceiver is further configured to send the first model parameter to the server.

In a seventh aspect, the present invention further provides a server, including a memory, a processor, and a program stored in the memory and executable on the processor; the processor implements the steps of any of the above federated learning-based model update methods when executing the program.

In an eighth aspect, the present invention further provides a terminal, including a memory, a processor, and a program stored in the memory and executable on the processor; the processor implements the steps of any of the above-described information transmission methods when executing the program.

In a ninth aspect, the present invention further provides a readable storage medium, on which a program is stored, which when executed by a processor, implements the steps in any of the above-described federal learning based model update methods or implements the steps in any of the above-described information transmission methods.

The technical scheme of the invention has the following beneficial effects:

in the embodiment of the invention, the server fuses the updated model parameters (namely the first model parameters) of each terminal according to the quality evaluation result of the data used by each terminal for updating the federal learning model to obtain the global model updating parameters (namely the second model parameters), and updates the model according to the global model updating parameters, so that the model quality and the modeling efficiency can be improved.

Drawings

Fig. 1 is a schematic flow chart of a model updating method based on federal learning according to a first embodiment of the present invention;

fig. 2 is a flowchart illustrating an information sending method according to a second embodiment of the present invention;

FIG. 3 is a schematic flow diagram of federated learning terminal selection and model update based on terminal-based training data quality assessment;

fig. 4 is a schematic structural diagram of a server according to a third embodiment of the present invention;

fig. 5 is a schematic structural diagram of a terminal according to a fourth embodiment of the present invention;

fig. 6 is a schematic structural diagram of a server according to a fifth embodiment of the present invention;

fig. 7 is a schematic structural diagram of a terminal according to a sixth embodiment of the present invention;

fig. 8 is a schematic structural diagram of a server in a seventh embodiment of the present invention;

fig. 9 is a schematic structural diagram of a terminal in an eighth embodiment of the present invention.

Detailed Description

In order to make the objects, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the drawings of the embodiments of the present invention. It is to be understood that the embodiments described are only a few embodiments of the present invention, and not all embodiments. All other embodiments, which can be derived by a person skilled in the art from the described embodiments of the invention, are within the scope of the invention.

Aiming at the phenomenon of low quality of a federal learning model, the inventor finds that the phenomenon is mainly caused by low quality of data participating in modeling. The global model parameters are fused from the parameters trained by each terminal, the quality of each terminal model directly influences the quality of the global model, and the quality of the data used for modeling by each terminal determines the quality of the model on the premise of determining the initial parameters and the model result.

In view of the problem of low modeling efficiency of the federal learning model, the inventor finds that the problem is mainly caused by the quality problem of the data sets used by the terminals participating in modeling, and particularly, in many cases, the quality of the data sets contained in the terminals participating in modeling is not uniform, the data of some terminals is defective, for example, the data amount is too small, the data loss is serious, the distribution is skewed, and the model convergence rate is reduced due to model parameters contributed to the central server by such terminals.

Therefore, the embodiment of the present invention optimizes the model fusion strategy starting from the data quality used by each terminal to update the model, which is described in detail below.

Referring to fig. 1, fig. 1 is a schematic flow chart of a model updating method based on federal learning according to an embodiment of the present invention, where the method is applied to a server, and includes the following steps:

step 11: obtaining first model parameters sent by at least two target terminals, and obtaining quality evaluation results of training data of the at least two target terminals, wherein the first model parameters are updated model parameters obtained by the target terminals performing model updating on a target federal learning model based on the training data;

specifically, each target terminal sends a first model parameter, each target terminal performs model update on the target federal learning model by using training data of each target terminal to obtain a corresponding first model parameter, and each training data has a corresponding quality evaluation result.

Wherein the target federated learning model may be a lateral federated learning model.

Step 12: according to the quality evaluation results of the training data of the at least two target terminals, carrying out weighted summation on the first model parameters to obtain second model parameters;

specifically, according to the quality evaluation result of the training data of each of the at least two target terminals, a weight value corresponding to each first model parameter is determined, and then each first model parameter is subjected to weighted summation based on the determined weight values. Further, the better the quality evaluation result of the training data is, for example, the higher the quality evaluation score is, the larger the weight value corresponding to the first model parameter of the target terminal may be.

Step 13: and updating the model of the target federal learning model according to the second model parameter. That is, the central server updates the global model according to the quality evaluation result of each target terminal for model training.

Wherein, the server is also a central server.

Aiming at the problem of low model quality, the following solutions exist at present: the model parameters of each terminal are arithmetically averaged. However, since the training data quality of each terminal is different, the contribution of the model parameters obtained by optimizing the model according to the training data to the global model is different, and therefore, directly performing arithmetic averaging on the model parameters of each terminal cannot guarantee effective optimization of the global model.

According to the method for updating the model based on the federal learning provided by the embodiment of the invention, the server fuses the updated model parameters (namely the first model parameters) of each terminal according to the quality evaluation result of the data used by each terminal for updating the federal learning model to obtain the updated global model parameters (namely the second model parameters), and updates the model according to the updated global model parameters, so that the quality of the model and the modeling efficiency can be improved.

The above-described model update method based on federal learning is exemplified below.

In an optional specific embodiment, the quality evaluation result is obtained by evaluating at least one of a data amount, data integrity and data distribution balance of the training data.

The real value of each terminal model is closely related to the quality of data used for training, and only the model trained by the high-quality data is meaningful. Therefore, quality assessment of data used for modeling is an important component of model value assessment. The data quality evaluation can be mainly carried out from the following aspects:

1. data volume: the larger the amount of data, the higher the value of the data for model training. The server issues basic data quantity values and gives a scoring rule, and each terminal gives a data quantity score to the terminal according to the scoring rule.

2. Data integrity:

1) data class integrity: the more data categories and data sets contained in the terminal, the higher the modeling value. And the server issues a data category integrity scoring rule, and each terminal counts the data category of the data set of the terminal and scores the data category of the terminal.

2) Data missing: the more complete the data, the fewer missing values and the higher the data quality. The server gives out a data integrity scoring rule, and each terminal scores itself according to the rule.

3. Data distribution balance: the more evenly distributed the data of each category, the higher the value of the data to the model training. The server gives a data distribution balance scoring rule, and each terminal scores itself according to the rule.

And after the scores of the three aspects are added, obtaining a terminal data quality score, namely a data quality evaluation result.

In the embodiment of the invention, the quality evaluation of the data used for updating the model at the terminal side can be finished by the terminal, and after the evaluation at the terminal side is finished, the quality evaluation result only needs to be sent to the server. The corresponding overhead mainly includes two parts: the calculation cost of data quality evaluation and the communication cost of uploading the evaluation result. The data quantity uploaded by the quality evaluation result is small (mainly comprising the identification of the terminal and the data evaluation score), so that the related communication overhead is very small, and the overall modeling process is not adversely affected. The calculation cost of the evaluation is related to the evaluation method, and can be performed by adopting preset software at the terminal or other modes, the common data quality evaluation index only relates to some simple statistical calculations, and the calculation cost of the evaluation is not large.

That is, in the embodiment of the present invention, the central server selects each round of terminals participating in modeling according to the quality evaluation result of the training data of each terminal model. In addition, since the training data of the terminal generally changes over time, for example, the latest data that can be used for model training is acquired, so that the central server needs to dynamically adjust the terminal participating in optimization according to the quality evaluation result of the latest training data of each terminal every time the model is updated.

Currently, there are also communication problems in the implementation of federal learning. Specifically, when the number of terminals (also referred to as clients) participating in learning is large, the communication traffic of interaction between each terminal and the central server during the training process is very large, so that the central server becomes a bottleneck of the training, for example, when each terminal transmits a parameter update value (i.e., an updated model parameter, i.e., the first model parameter) to the central server, a large amount of traffic arrives at the central server at the same time, and a throughput collapse problem occurs. In the related art, in order to solve the communication problem in the modeling process, strategies of selecting a part of terminals to participate in modeling and performing communication resource allocation according to the specific conditions of the terminals are mainly adopted. Since the modeling process is a process of multiple iterations, and the training data held by each terminal is often time-varying, the terminals participating in the model update in the current round are usually updated at each model update. The following methods are used to select the terminals participating in the model update:

1. randomly selecting a terminal according to a certain proportion;

2. calculating the time for completing model updating and uploading according to the data volume, computing resources, communication resources and the like of each terminal, and selecting the terminal capable of completing updating and uploading in the given time;

3. and evaluating the updated gradient of each terminal, and allocating communication resources to the terminal according to the contribution degree of the gradient to the model.

The method mainly solves the problem of throughput collapse from the aspects of reducing communication traffic and reasonably configuring communication resources. The method for completely randomly selecting part of the terminals reduces communication traffic, but the randomly selected terminals probably enable part of the terminals not to have the opportunity to participate in model optimization, so that the generalization capability of the model is limited; the strategy of selecting the terminal capable of completing updating and uploading in the given time solves the problem of excessive number of terminals participating in optimizing the model on one hand, and can ensure that the modeling time of the global model is not too long on the other hand, thereby improving the modeling efficiency, but possibly leading part of the terminals not to be selected all the time and not to participate in model optimization due to factors such as data volume, communication resources and the like, thereby influencing the model quality; according to the contribution degree of the gradient to the model, the method for allocating the communication resources to the terminal is equivalent to improving the modeling efficiency of the global model, but the gradient is only used for measuring the contribution of the terminal to the model optimization, so that the method is not accurate enough, the situation that the gradient performance is good, but the data distribution used for training by the terminal is biased or the data volume is small can occur, and the quality of the model is influenced.

Specifically, after the server obtains the quality evaluation result of the training data which can be used by the terminal for model updating, the server may select, according to a certain proportion, a part of terminals with higher scores as the target terminals to participate in the model optimization of the current round. The method comprises the steps that a server sends model initial parameters to a selected terminal (namely a target terminal), the terminal starts model training, and updated model parameters are encrypted and sent to the server after the training is finished.

The embodiment of the invention provides a method for selecting the terminal participating in model optimization according to the quality of training data used by the terminal for updating the model, which not only can reduce communication traffic, but also can improve the quality and modeling efficiency of a federal learning model.

Specifically, after the server selects the target terminal from the terminals according to the quality evaluation results of the training data of the at least two terminals, the quality evaluation result of the training data of the target terminal can be screened from the quality evaluation results of the training data of the at least two terminals.

Specifically, when the target terminal is selected from the terminals according to the quality evaluation results of the training data of the at least two terminals, a certain number of terminals with higher quality evaluation scores may be selected as the target terminals according to the current communication conditions.

the terminal needs to update the target federal learning model according to the initial parameters of the target federal learning model. Therefore, the terminal starts to update the target federated learning model only after receiving the initial parameters of the target federated learning model again. In other words, the initial parameters received into the target federated learning model are an indication that the terminal starts to perform the model update of the current round.

And taking the first terminal corresponding to the first model parameter received in the preset time as the target terminal. Then, the quality evaluation result of the training data of the target terminal can be selected from the quality evaluation results of the training data of the at least two terminals, and the weight value of the first model parameter of each target terminal is determined according to the quality evaluation result of the training data of the target terminal. And finally, carrying out weighted summation on the first model parameters of the target terminals according to the weight values of the first model parameters of the target terminals to obtain second model parameters.

In the embodiment of the invention, the server can set a time limit (namely a preset time period), the updated model parameters of each terminal received in the time limit are fused, and the updated model parameters are weighted and summed according to the quality evaluation score of each terminal data to update the global model. And for the model parameters of the terminal which cannot be received after exceeding the time limit, not participating in the overall model updating.

In the embodiment of the invention, the quality evaluation result of each terminal data is taken as the main, each terminal completes one round of model updating time as the auxiliary, and the terminal participating in each round of model updating is selected.

In addition, the server may broadcast, to all terminals, an Identifier (ID) of a terminal participating in model update in the current round, and declare that the current round of model update is completed, while or after performing model update on the target federal learning model according to the second model parameter, by performing weighted summation on the first model parameter according to the quality evaluation result of the training data of the at least two target terminals to obtain the second model parameter.

The embodiment of the invention has the main advantages that the quality evaluation result of the training data is incorporated into the terminal selection method and the central server model updating strategy participating in model optimization, the communication problem of federal learning is solved (namely, the communication load is reduced), the model quality of each terminal participating in model updating can be ensured, the global model quality of the central server is further ensured, and the modeling efficiency is improved.

Referring to fig. 2, fig. 2 is a flowchart illustrating an information sending method according to a second embodiment of the present invention, where the method is applied to a terminal, and includes the following steps:

step 21: sending a quality evaluation result of a first model parameter and training data to a server, so that the server performs weighted summation on the first model parameters sent by at least two terminals according to the quality evaluation result to obtain a second model parameter, and performing model updating on a target federated learning model according to the second model parameter;

Specifically, each terminal sends a first model parameter, each terminal performs model update on the target federal learning model by using training data of each terminal to obtain a corresponding first model parameter, and the training data of each terminal has a corresponding quality evaluation result. The server is also a central server.

2. Data integrity:

In the embodiment of the invention, the quality evaluation of the data used for updating the model at the terminal side can be finished by the terminal, after the evaluation at the terminal side is finished, only the quality evaluation result needs to be sent to the server, and the data volume uploaded by the quality evaluation result is very small (mainly comprising the identification and the data evaluation score of the terminal), so that the related communication overhead is very small, and the adverse effect on the whole modeling process can be avoided. The specific quality evaluation process can be performed by terminal preset software or other modes, the commonly used data quality evaluation indexes only relate to some simple statistical calculations, and the evaluation calculation cost is not large.

In addition, before sending the first model parameter and the quality evaluation result of the training data to the server, the method further includes:

receiving an evaluation rule sent by a server;

and according to the evaluation rule, carrying out quality evaluation on the training data to obtain the quality evaluation result.

Optionally, before sending the first model parameter and the quality evaluation result of the training data to the server, the method further includes:

and determining whether to participate in the model updating according to the data volume and/or the data integrity of the training data.

Specifically, if the data volume is less than or equal to the preset threshold, the model is not updated. And if the number of the classes contained in the training data is less than half of the total number of the classes, not participating in the model updating. And if the data type of the missing data exceeding the first proportion threshold value exceeds a second proportion threshold value of the total type, not participating in the model updating.

sending a quality evaluation result of the training data to the server;

specifically, the server receives quality evaluation results of training data sent by a plurality of terminals, and after receiving the quality evaluation results of the training data, screens out target terminals from the terminals according to the quality evaluation results of the training data of the terminals, for example, sorting scores of the quality evaluation results of the training data of the terminals in descending order, and selecting a certain number of terminals with the scores of the quality evaluation results before as the target terminals. And then sending the initial parameters of the target federated learning model to the screened terminals.

And sending the initial parameters of the target federated learning model to a terminal, wherein the initial parameters of the target federated learning model indicate that the terminal is a terminal selected to participate in model updating. After receiving the initial parameters of the target federated learning model, the terminal may begin model update training.

and sending the first model parameters to the server.

Currently, there are also communication problems in the implementation of federal learning. Specifically, when the number of terminals (also referred to as clients) participating in learning is large, the communication traffic of interaction between each terminal and the central server during the training process is very large, so that the central server becomes a bottleneck of the training, for example, when each terminal transmits a parameter update value (i.e., an updated model parameter, i.e., the first model parameter) to the central server, a large amount of traffic arrives at the central server at the same time, and a throughput collapse problem occurs.

Specifically, after each terminal sends the quality evaluation result of the data that can be used for model updating to the server, the server may select, according to a certain proportion, a part of terminals with higher scores as the target terminals to participate in the model optimization of the current round. The method comprises the steps that a server sends model initial parameters to a selected terminal (namely a target terminal), the terminal starts model training, and updated model parameters are encrypted and sent to the server after the training is finished.

The embodiment of the invention provides a method for selecting the terminal participating in model optimization according to the quality of data used by the terminal for updating the model, which not only can reduce communication traffic, but also can improve the quality and modeling efficiency of a federal learning model.

Referring to fig. 3, fig. 3 is a schematic overall flow chart of a model updating method based on federal learning according to an embodiment of the present invention, where the overall flow of model updating is specifically as follows:

1. the central server broadcasts a new round of model updating start to all the terminals;

2. each terminal calculates the quality evaluation score of the data which can be used for model updating, namely each terminal quantitatively evaluates the quality of the data which is used for model training;

3. each terminal sends the quality evaluation score calculated in the previous step to a central server;

4. specifically, the central server may send initial parameters (which may also be referred to as model initial values) of the model to a certain number of terminals with quality evaluation scores of data arranged in front (ordered from large to small), and start the current round of model updating;

5. each terminal updates the model and sends the updated model parameters (which can be encrypted) to the central server after the local model is updated;

6. and the central server performs weighted summation on the model parameters of each terminal according to the quality evaluation score of each terminal data, and updates the global model.

The embodiment of the invention provides a method which takes the data quality evaluation of each terminal as a main basis from two aspects of selection of the terminals participating in model updating and a strategy that a central server fuses parameters of each terminal model when each round of model updating.

The method for updating the model based on federal learning is described below by taking the construction of a home care rehabilitation management model for stroke patients as an example.

Rehabilitation therapy is an important link in the treatment process of stroke patients, and means that the stroke patients are treated and rehabilitated under the guidance of medical care personnel, the treatment and rehabilitation are often carried out in rehabilitation hospitals, and the medical care personnel need to nurse and treat the patients from the aspects of disease condition monitoring, medication, nursing, diet, daily life, rehabilitation and the like. Due to limited medical resources, all patients cannot be guaranteed to be served by specialized doctors in the rehabilitation hospital. Therefore, a cerebral apoplexy patient rehabilitation management model needs to be constructed so that the community general practitioners can care and treat the patients with assistance of family members under the guidance of the model.

The model needs a large amount of data of rehabilitation hospital patients, the data of the patients belongs to individual privacy and cannot be provided for institutions except hospitals, and the data of one hospital is far insufficient for constructing the model, so that a federal learning method needs to be applied to jointly construct a home maintenance rehabilitation model for stroke patients.

The data involved in modeling includes two broad categories: patient personal status data including patient demographic data, past medical history data, vital sign data (e.g., blood pressure, blood glucose, heart rate, etc.), daily data (e.g., sleep status), etc.; the hospital treatment related data comprises medication, rehabilitation exercise schemes and the like, and can be summarized and classified into a plurality of treatment rehabilitation schemes after being combined. The constructed model takes the personal state data of the patient as input and the treatment and rehabilitation scheme as output.

Due to differences in the scale, professional level, informatization level, regions and the like of rehabilitation hospitals, data quality of each hospital is different, and in order to guarantee model quality and modeling efficiency, data quality of each hospital needs to be evaluated. The specific method comprises the following steps:

in the modeling data, the treatment and rehabilitation schemes are C-type, the input personal state data of the patient is D-dimensional, and the constructed model parameters are P-type. The total evaluation of the data quality of each terminal is 100 points, wherein the data volume and the data integrity respectively account for 40 points, and the data distribution balance accounts for 20 points.

1. Data volume: the basic data volume is determined to be N-10P, and 32 points are obtained when the basic data volume is met; every time the number of the above-mentioned pieces exceeds P, the score is increased by 1 point; and when the data volume is less than or equal to 2P (including 2P), the terminal does not participate in the modeling.

2. Data integrity:

the data set comprises all D category data, and 20 points are obtained; when the number of the categories is reduced by 4 points every time [0.1D ] categories are reduced, and the number of the included categories is less than half of the total number of the categories, the terminal does not participate in the modeling;

(II) no data is lost, and 20 points are obtained; the dimension with the deletion value accounting for more than 10% is called as the dimension with deletion, and the score is reduced by 1 when the dimension with deletion accounts for 10% of the total amount of the features; the score is reduced by 1 point when the missing data quantity of the 'missing dimensionality' is increased by 10 percent; the 'missing dimension' of more than 50% of missing data exceeds 20% of the total characteristic number, and the terminal does not participate in the modeling.

3. Data distribution balance: and calculating the ratio of each category of data in the total data, and further calculating the entropy E of data distribution, wherein 20E is the score.

And adding the three scores to obtain the evaluation score of each terminal data.

The model construction process is as follows:

1. each hospital information center collects data required by modeling;

2. the central server sends data quality evaluation software to each hospital information center, and the information centers are provided with the evaluation software;

3. the central server broadcasts the information of the start of modeling to each hospital;

4. each hospital evaluates the data quality thereof, obtains a data quality score and sends the data quality score to a central server;

5. the central server determines hospitals with scores higher than 50 points to participate in modeling according to the received data quality scores, and sends model initial values to the hospitals;

6. the hospitals participating in modeling use the data of the hospitals to update the models, encrypt the updated model parameters and send the encrypted model parameters to the central server;

7. the central server sets a time threshold value, and generates an integral model by using the hospital model parameters received in the time threshold value. Specifically, the hospital parameters are weighted and summed according to the evaluation scores of the hospital data. Model parameters received after the time threshold is exceeded do not participate in the updating of the model in the current round;

and the central server broadcasts the completion of the modeling in the current round.

Referring to fig. 4, fig. 4 is a schematic structural diagram of a server according to a third embodiment of the present invention, where the server 40 includes:

an obtaining module 41, configured to obtain first model parameters sent by at least two target terminals, and obtain quality evaluation results of training data of the at least two target terminals, where the first model parameters are updated model parameters obtained by the target terminals performing model update on a target federated learning model based on the training data;

the weighted summation module 42 is configured to perform weighted summation on the first model parameter according to the quality evaluation results of the training data of the at least two target terminals to obtain a second model parameter;

and the global model updating module 43 is configured to perform model updating on the target federal learning model according to the second model parameter.

Optionally, the server 40 further includes:

Optionally, the screening module includes:

The embodiment of the present invention is a product embodiment corresponding to the above method embodiment, and therefore, detailed description is omitted here, and please refer to the first embodiment in detail.

Referring to fig. 5, fig. 5 is a schematic structural diagram of a terminal according to a fourth embodiment of the present invention, where the terminal 50 includes:

a sending module 51, configured to send a quality evaluation result of the first model parameter and the training data to a server, so that the server performs weighted summation on the first model parameters sent by at least two terminals according to the quality evaluation result to obtain a second model parameter, and performs model update on a target federal learning model according to the second model parameter;

Optionally, the sending module 51 includes:

The embodiment of the present invention is a product embodiment corresponding to the above method embodiment, and therefore, detailed description is omitted here, and please refer to the second embodiment.

Referring to fig. 6, fig. 6 is a schematic structural diagram of a server according to a fifth embodiment of the present invention, where the server 60 includes: a transceiver 61 and a processor 62;

the transceiver 61 is configured to obtain first model parameters sent by at least two target terminals, and obtain quality evaluation results of training data of the at least two target terminals, where the first model parameters are updated model parameters obtained by the target terminals performing model update on a target federated learning model based on the training data;

the processor 62 is configured to perform weighted summation on the first model parameter according to the quality evaluation results of the training data of the at least two target terminals to obtain a second model parameter;

Optionally, the transceiver 61 is configured to obtain quality evaluation results of training data of at least two terminals;

the processor 62 is configured to select the target terminal from the terminals according to the quality evaluation result of the training data of the at least two terminals.

Optionally, the processor 62 is configured to select a first terminal from the terminals according to a quality evaluation result of the training data of the at least two terminals;

the transceiver 61 is configured to send the initial parameters of the target federated learning model to the first terminal;

the transceiver 61 is further configured to receive, within a preset time period, a first model parameter sent by the first terminal, where the first model parameter is an updated model parameter obtained by the first terminal performing model update on the target federal learning model after receiving the initial parameter;

the processor 62 is further configured to use the first terminal corresponding to the first model parameter received within the preset time as the target terminal.

Referring to fig. 7, fig. 7 is a schematic structural diagram of a terminal according to a sixth embodiment of the present invention, where the terminal 70 includes: a transceiver 71 and a processor 72;

the transceiver 71 is configured to send a quality evaluation result of the first model parameters and the training data to a server, so that the server performs weighted summation on the first model parameters sent by at least two terminals according to the quality evaluation result to obtain second model parameters, and performs model update on a target federal learning model according to the second model parameters;

Optionally, the transceiver 71 is configured to send a quality evaluation result of the training data to the server;

the transceiver 71 is further configured to receive initial parameters of the target federal learning model sent by the server after receiving the quality evaluation result of the training data;

the processor 72 is configured to perform model update on the target federal learning model according to the initial parameters and the training data, so as to obtain first model parameters;

the transceiver 71 is further configured to send the first model parameters to the server.

Referring to fig. 8, fig. 8 is a schematic structural diagram of a server according to a seventh embodiment of the present invention, where the server 80 includes a processor 81, a memory 82, and a program stored in the memory 82 and capable of running on the processor 81; the processor 81 implements the following steps when executing the program:

Optionally, when the processor 81 executes the program, the following steps may be further implemented:

before the obtaining of the first model parameters sent by the at least two target terminals, the method further includes:

the selecting the target terminal from the terminals according to the quality evaluation results of the training data of the at least two terminals includes:

The specific working process of the embodiment of the present invention is the same as that of the first embodiment of the method, and therefore, detailed description is not repeated here, and please refer to the description of the method steps in the first embodiment.

Referring to fig. 9, fig. 9 is a schematic structural diagram of a terminal according to an eighth embodiment of the present invention, where the terminal 90 includes a processor 91, a memory 92, and a program stored in the memory 92 and capable of running on the processor 91; the processor 91 implements the following steps when executing the program:

Optionally, when the processor 91 executes the program, the following steps may be further implemented:

the sending of the quality evaluation result of the first model parameter and the training data to the server includes:

sending a quality evaluation result of the training data to the server;

and sending the first model parameters to the server.

The specific working process of the embodiment of the present invention is the same as that of the second embodiment of the method, and therefore, the detailed description thereof is omitted, and refer to the description of the method steps in the second embodiment.

An embodiment ninth of the present invention provides a readable storage medium, where a program is stored, and the program, when executed by a processor, implements the steps in any one of the federal learning based model update methods in the first embodiment or the steps in any one of the information sending methods in the second embodiment. Please refer to the above description of the method steps in the corresponding embodiments.

The terminal in the embodiments of the present invention may be a wireless terminal or a wired terminal, and the wireless terminal may be a device providing voice and/or other service data connectivity to a user, a handheld device having a wireless connection function, or other processing devices connected to a wireless modem. A wireless terminal, which may be a mobile terminal such as a mobile telephone (or "cellular" telephone) and a computer having a mobile terminal, e.g., a portable, pocket, hand-held, computer-included, or vehicle-mounted mobile device, may communicate with one or more core networks via a Radio Access Network (RAN), and may exchange language and/or data with the RAN. For example, devices such as Personal Communication Service (PCS) phones, cordless phones, Session Initiation Protocol (SIP) phones, Wireless Local Loop (WLL) stations, and Personal Digital Assistants (PDAs) are used. A wireless Terminal may also be referred to as a system, a Subscriber Unit (Subscriber Unit), a Subscriber Station (Subscriber Station), a Mobile Station (Mobile), a Remote Station (Remote Station), a Remote Terminal (Remote Terminal), an Access Terminal (Access Terminal), a User Terminal (User Terminal), a User Agent (User Agent), and a Terminal (User Device or User Equipment), which are not limited herein.

The readable storage medium includes a computer readable storage medium. Computer-readable storage media, including both non-transitory and non-transitory, removable and non-removable media, may implement information storage by any method or technology. The information may be computer readable instructions, data structures, modules of a program, or other data. Examples of storage media include, but are not limited to, phase change memory (PRAM), Static Random Access Memory (SRAM), Dynamic Random Access Memory (DRAM), other types of Random Access Memory (RAM), Read Only Memory (ROM), Electrically Erasable Programmable Read Only Memory (EEPROM), flash memory or other memory technology, compact disc read only memory (CD-ROM), Digital Versatile Discs (DVD) or other optical storage, magnetic cassettes, magnetic tape storage or other magnetic storage devices, or any other non-transmission medium that can be used to store information that can be accessed by a computing device.

While the foregoing is directed to the preferred embodiment of the present invention, it will be understood by those skilled in the art that various changes and modifications may be made without departing from the spirit and scope of the invention as defined in the appended claims.

Claims

1. A model updating method based on federal learning is applied to a server and is characterized by comprising the following steps:

2. The method of claim 1, wherein the quality assessment result is obtained by assessing at least one of data quantity, data integrity and data distribution balance of the training data.

3. The method of claim 1, wherein the target federal learning model is a stroke patient home care rehabilitation management model.

4. The method of claim 1, wherein before obtaining the first model parameters sent by the at least two target terminals, the method further comprises:

5. The method according to claim 4, wherein the selecting the target terminal from the terminals according to the quality evaluation results of the training data of the at least two terminals comprises:

6. An information sending method applied to a terminal is characterized by comprising the following steps:

7. The method of claim 6, wherein the quality assessment result is obtained by assessing at least one of data quantity, data integrity and data distribution balance of the training data.

8. The method of claim 6, wherein the target federal learning model is a stroke patient home care rehabilitation management model.

9. The method of claim 6, wherein sending the quality assessment results of the first model parameters and the training data to the server comprises:

sending a quality evaluation result of the training data to the server;

and sending the first model parameters to the server.

10. A server, comprising:

11. A terminal, comprising:

12. A server, comprising: a transceiver and a processor;

13. A terminal, comprising: a transceiver and a processor;

14. A server comprising a memory, a processor, and a program stored on the memory and executable on the processor; wherein the processor, when executing the program, implements the steps in the federated learning-based model update method of any one of claims 1-5.

15. A terminal comprising a memory, a processor, and a program stored on the memory and executable on the processor; characterized in that the processor implements the steps in the method for sending information according to any one of claims 6 to 9 when executing the program.

16. A readable storage medium on which a program is stored, the program, when executed by a processor, implementing the steps in the federal learning based model update method as claimed in any one of claims 1 to 5 or implementing the steps in the information transmitting method as claimed in any one of claims 6 to 9.