CN116719607A - Model updating method and system based on federal learning - Google Patents

Abstract

The invention provides a model updating method and system based on federal learning, wherein the method comprises the following steps: updating clusters to which the target client belongs according to synthetic data of the target client participating in federal learning until the initialized K clusters reach a stable distribution state, and sending a target cluster model in each cluster to the target client, wherein the synthetic data is obtained according to the received data distribution information of the target client, the stable distribution state is that the cluster to which the target client belongs is not changed any more, and the target cluster model is determined according to the cluster model in each cluster in which the initialized K clusters reach the stable distribution state; a first loop process is performed until the candidate models in the target client converge. The invention can fully utilize the data of all clients in the data heterogeneous environment, and provides an optimal personalized model for each client under the conditions of limited network communication resources, limited computing resources and limited storage resources of the clients.

Description

Model update method and system based on federated learning

Technical field

The present invention relates to the technical field of federated learning, and in particular to a model update method and system based on federated learning.

Background technique

As edge devices become more common in modern society, distributed private data is growing rapidly. This data offers huge opportunities for artificial intelligence applications. However, these data are stored in silos on end users’ own devices, such as mobile phones and personal computers, and are mostly highly sensitive. With the introduction of data privacy protection laws, such as the General Data Protection Regulation (GDPR), there is an increasing need to protect privacy in artificial intelligence, so this data is usually not made public. In recent years, Artificial Intelligence (AI) is still developing rapidly, followed by increasingly larger AI models, and these models also require an increasing amount of training data. In order to utilize data from multiple institutions or individuals and achieve information fusion while protecting data privacy, federated learning has become an effective method to train AI models. This approach allows leveraging large amounts of data for model training while maintaining data confidentiality.

Federated Learning (FL) is a machine learning paradigm proposed in recent years. It adopts a "client-server" architecture model and aims to solve the above problems. Figure 1 is the architecture of federated learning provided by existing technology. Schematic diagram, as shown in Figure 1. The federated learning training model iterates through three steps until convergence: (1) The server sends the global model through the network to the clients participating in this round of training; (2) After the client receives the global model, it overwrites the old local model and then combines the client's local Private data for training; (3) After the training is completed, the client uploads the trained local model to the server through the network; the server performs model aggregation after receiving the trained local model uploaded by the client that meets the set number, and generates New global model. It can be seen that federated learning promotes collaborative training among customers in multiple data warehouses in a way that protects privacy. During the training process, customers' private data is stored locally and is not shared directly, so that customers can achieve comparison while protecting privacy. Works alone for better performance.

However, since the data sources and computing nodes are the end-user's personal devices, each customer has his or her own independent data, and there is no Independent and Identically Distributed (IID) training data, that is, the private data of the customers involved in training. The scale and distribution may be different, which leads to the single global model generated by traditional federated learning not performing well for some customers. It is even worse than simply using their own local data (even if it is small) to train a local model. model effect.

Personalized federated learning aims to use the federated learning process to make up for the problem of insufficient local data on the client, learn a model suitable for local data sets for each client, and improve the training effect of the personalized federated learning model.

In the field of personalized federated learning, some federated learning methods try to solve this problem by performing local fine-tuning after the global model is trained. However, the personalized performance of this method still depends on the generalization performance of the global model and does not solve the essential problem. Other federated learning algorithms explore distributed clustering technology to better simulate non-IID data and use multiple iterative clustering schemes, but this method is very unfriendly to clients with limited communication bandwidth and computing resources.

Therefore, it is possible to make full use of the data of all customers in a heterogeneous data environment, and learn an optimal personalized model for each client under the condition of limited network communication resources, limited computing resources and limited storage resources of the client, to achieve Higher performance is one of the key issues of federated learning technology.

Contents of the invention

The model update method and system based on federated learning provided by the present invention are used to solve the above problems existing in the existing technology.

The invention provides a model update method based on federated learning, which is applied to servers participating in federated learning, including:

According to the synthetic data of the target clients participating in the federated learning, the cluster to which the target client belongs is updated until the initialized K clusters reach a stable distribution state, and the target clustering model in each cluster is sent to the Target client, the synthetic data is obtained based on the received data distribution information of the target client, the stable distribution state means that the cluster to which the target client belongs does not change, and the target client includes N clients end, the target clustering model is determined based on the clustering model in each cluster when the initialized K clusters reach a stable distribution state;

Execute the following first loop process until the candidate model in the target client converges, the candidate model is obtained after the target client updates the received target clustering model in each cluster;

The first cycle process includes:

Receive the candidate model sent by the target client and update the candidate model;

The updated candidate model is sent to the target client, so that the target client updates the candidate model.

According to a model update method based on federated learning provided by the present invention, which is applied to servers participating in federated learning, updating the cluster to which the target client belongs based on the synthetic data of the target client includes:

Initialize the cluster to which the target client belongs;

According to the empirical loss function of the clustering model in each cluster, obtain the empirical loss value of the clustering model in each cluster for the synthetic data of each client;

According to the empirical loss value of the clustering model of the synthetic data of each client in each cluster, the cluster to which the target client belongs is updated to determine when the K clusters reach a stable distribution state. The clients included in each cluster are the clients among the target clients that satisfy the experience loss value of the clustering model of the synthetic data in each cluster reaching the minimum value.

According to a model update method based on federated learning provided by the present invention, it is applied to servers participating in federated learning. According to the empirical loss value of the clustering model of the synthetic data of each client in each cluster, the Update the cluster to which the target client belongs, including:

Execute the following second loop process until the K clusters reach a stable distribution state;

The second cycle process includes:

Update the clustering model within each cluster based on the gradient descent method;

Based on the synthetic data of the clients included in each cluster, update the empirical loss value of the updated clustering model of the clients included in each cluster within the cluster;

The cluster to which the target client belongs is updated according to the empirical loss value of the updated clustering model of the clients contained in each cluster in the cluster.

According to a model update method based on federated learning provided by the present invention, which is applied to servers participating in federated learning, receiving the candidate model sent by the target client and updating the candidate model includes:

The optimization objective function of the candidate model is determined according to the sum of the first optimization objective function and the second optimization objective function. The first optimization objective function is the sum of the loss functions of the candidate models of the target client. The second optimization objective function is the sum of the loss functions of the candidate models of the target client. The optimization objective function is the sum of nonlinear functions of differences between the target clustering model received by any client and the target clustering model received by the remaining clients in the target client;

Optimize the first optimization objective function based on the gradient descent method;

Optimize the second optimization objective function based on the approximate point method;

The candidate model is updated according to the candidate model when the optimization objective function takes the minimum value.

The present invention also provides a model update method based on federated learning, which is applied to target clients participating in federated learning, including:

Send the data distribution information of the target client to the server participating in the federated learning, so that the server updates the cluster to which the target client belongs based on the synthetic data of the target client, where the target client includes N clients, the synthetic data is obtained according to the received data distribution information of the target client;

When the server determines that the initialized K clusters have reached a stable distribution state, receive the target clustering model in each cluster sent by the server, and the stable distribution state means that the cluster to which the target client belongs does not change. , the target clustering model is determined based on the clustering model in each cluster when the initialized K clusters reach a stable distribution state;

Execute the following first loop process until the candidate model converges, the candidate model is obtained after the target client updates the received target clustering model in each cluster;

The first cycle process includes:

Send the candidate model to the server so that the server updates the received candidate model;

Receive the updated candidate model sent by the server, and update the updated candidate model.

According to a model update method based on federated learning provided by the present invention, it is applied to a target client participating in federated learning. The method for obtaining the data distribution information of the target client includes:

Based on knowledge distillation technology, pre-train teacher network;

Use the trained teacher network as the discriminator of a generative adversarial network, which is deployed on the target client;

Train the generator in the generative adversarial network based on the discriminator until the value of the loss function of the generative adversarial network is less than a preset value;

According to the trained generator, the data distribution information of the target client is determined.

The present invention also provides a model update system based on federated learning, applied to servers participating in federated learning, including: a first sending module and a first update module;

The first sending module is used to update the cluster to which the target client belongs based on the synthetic data of the target client participating in the federated learning until the initialized K clusters reach a stable distribution state, and update the clusters within each cluster. The target clustering model is sent to the target client, the synthetic data is obtained according to the received data distribution information of the target client, and the stable distribution state means that the cluster to which the target client belongs does not change, The target clients include N clients, and the target clustering model is determined based on the clustering model in each cluster when the initialized K clusters reach a stable distribution state;

The first update module is configured to perform the following first loop process until the candidate model in the target client converges. The candidate model is the target cluster within each cluster received by the target client. Obtained after the model is updated;

The first cycle process includes:

Receive the candidate model sent by the target client and update the candidate model;

The second sending module is used to send the updated candidate model to the target client, so that the target client updates the candidate model.

The present invention also provides a model update system based on federated learning, applied to target clients participating in federated learning, including: a second sending module, a receiving module and a second update module;

The second sending module is used to send the data distribution information of the target client to the server participating in the federated learning, so that the server can analyze the cluster to which the target client belongs based on the synthetic data of the target client. Update is performed, the target client includes N clients, and the synthetic data is obtained according to the received data distribution information of the target client;

The receiving module is configured to receive the target clustering model in each cluster sent by the server when the server determines that the initialized K clusters have reached a stable distribution state, and the stable distribution state is the target The cluster to which the client belongs no longer changes, and the target clustering model is determined based on the clustering model in each cluster when the initialized K clusters reach a stable distribution state;

The second update module is used to perform the following first loop process until the candidate model converges. The candidate model is obtained after the target client updates the received target clustering model in each cluster;

The first cycle process includes:

Send the candidate model to the server so that the server updates the received candidate model;

The second receiving module is configured to receive the updated candidate model sent by the server, and update the updated candidate model.

The present invention also provides an electronic device, including a processor and a memory storing a computer program. When the processor executes the program, it implements any one of the above federated learning-based model update methods.

The present invention also provides a non-transitory computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, it implements any of the above federated learning-based model update methods.

The present invention also provides a computer program product, which includes a computer program. When the computer program is executed by a processor, the computer program implements any one of the above federated learning-based model updating methods.

In the federated learning-based model update method and system provided by the present invention, target clients participating in collaborative training upload data distribution information to the server. The server clusters and divides the target clients based on the uploaded data distribution information to obtain the target clients. The cluster to which the client belongs saves the computing resources and storage resources of the target client, reduces the communication overhead between the target client and the server, and learns an optimal personalized model (i.e., a converged candidate model) for each client.

Description of the drawings

In order to explain the present invention or the technical solutions in the prior art more clearly, the drawings needed to be used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings in the following description are of the present invention. For some embodiments of the invention, those of ordinary skill in the art can also obtain other drawings based on these drawings without exerting creative efforts.

Figure 1 is a schematic diagram of the architecture of federated learning provided by existing technology;

Figure 2 is one of the flow diagrams of the model update method based on federated learning provided by the present invention;

Figure 3 is a schematic flow chart of the server provided by the present invention updating the cluster to which the target client belongs;

Figure 4 is a schematic diagram of generating a personalized model provided by the present invention;

Figure 5 is the second schematic flow chart of the model update method based on federated learning provided by the present invention;

Figure 6 is a schematic diagram of the distribution of generated data with a generative adversarial network provided by the present invention;

Figure 7 is the third schematic flow chart of the federated learning model update method provided by the present invention;

Figure 8 is one of the structural schematic diagrams of the model update system based on federated learning provided by the present invention;

Figure 9 is the second structural schematic diagram of the model update system based on federated learning provided by the present invention;

Figure 10 is a schematic diagram of the physical structure of the electronic device provided by the present invention.

Detailed ways

In order to make the purpose, technical solutions and advantages of the present invention more clear, the technical solutions in the present invention will be clearly and completely described below in conjunction with the accompanying drawings of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention. , not all examples. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without making creative efforts fall within the scope of protection of the present invention.

The model update method based on federated learning provided by the present invention learns the data distribution of the client and completes the clustering work of the client on the server based on the data distribution, thereby reducing the communication overhead between the client and the server and saving the client's computing resources and Storage resources, as well as high-performance personalized models for learning clients within clustering clusters, solve the problem of the heterogeneous distribution of underlying client data in federated learning, which results in poor results in the global model of collaborative learning and slow convergence of model training. The problem includes the following steps: the client uses generative adversarial network technology to learn the different data distributions of the underlying customers, and sends these data distributions to the server through the communication channel; the server uses these data distribution information to perform similarity aggregation on the participating clients. Classes allow clients with similar distributions to be gathered together; within each cluster, the similarity between clients is fully utilized to collaboratively train a personalized model belonging to each client for one client. The present invention can use the data distribution information of the client to complete the clustering work on the server side, effectively reduce the communication load, improve communication efficiency, reduce communication bandwidth pressure, save the computing resources and storage resources of the local client, and significantly improve the performance of the personalized model. The specific effect is as follows:

Figure 2 is one of the flow diagrams of the model update method based on federated learning provided by the present invention. As shown in Figure 2, the method includes:

Step 110: Update the cluster to which the target client belongs based on the synthetic data of the target clients participating in the federated learning until the initialized K clusters reach a stable distribution state, and send the target clustering model in each cluster. For the target client, the synthetic data is obtained according to the received data distribution information of the target client, the stable distribution state means that the cluster to which the target client belongs does not change, and the target client includes N clients, the target clustering model is determined based on the clustering model in each cluster when the initialized K clusters reach a stable distribution state;

Step 120: Execute the following first loop process until the candidate model in the target client converges. The candidate model is obtained after the target client updates the received target clustering model in each cluster. ;

The first cycle process includes:

Receive the candidate model sent by the target client and update the candidate model;

The updated candidate model is sent to the target client, so that the target client updates the candidate model.

It should be noted that the execution subject of the above method can be a computer device or can be applied to a server participating in federated learning.

Optionally, the target client may specifically be N clients randomly selected by the server participating in federated learning from the clients participating in federated learning, where N is a positive integer.

The synthetic data of the N clients can be obtained specifically from the data distribution information of the N clients. The data distribution information is the generator in each client. The generator can be specifically obtained by initializing the data deployed on the target client. Obtained after training the local model (adversarial generative network).

The server receives the data publishing information of the N clients, performs clustering on the server side, and generates synthetic data corresponding to each client based on the data distribution information uploaded by different clients (i.e., the trained generator). For example, the server will The noisy data is input to the trained generator to generate synthetic data for each client.

The server receives the data distribution information G 1 G 2...G i...G n of the client C ₁ C ₂ ...C _i ...C _n numbered _{1, 2,} ..., _i ..., _N. , based on the data distribution information G ₁ G ₂ ...G _i ...G _n , generate synthetic data corresponding to each client. The synthetic data of the i-th client information on the server is recorded as D′ _i ; the synthetic data set The size is recorded as |D′ _i |.

And based on the synthetic data corresponding to the N clients, the cluster to which the target client belongs is updated.

The cluster can specifically be any cluster among the K clusters initialized by the server. Each cluster corresponds to a clustering model. The parameters of the clustering model are: Among them,/> Represents the parameters of the jth clustering model, K is a positive integer.

The server performs multiple iterations on the server side based on the synthetic data corresponding to the target client, and estimates the cluster to which each client belongs based on the minimum loss calculated by the empirical loss function of the clustering model in the K clusters until the initialized K clusters A stable distribution state is reached. The stable distribution state specifically means that the cluster to which each client in the target client belongs will no longer change, that is, the clients included in each cluster will no longer change.

After determining that the initialized K clusters have reached a stable distribution state, the server sends the target clustering model in each cluster to the clients in the cluster according to the cluster to which each of the N clients belongs.

After the target client receives the target clustering model sent by the server, it updates it using the gradient descent method.

The following first loop process is executed until the candidate model in the target client converges.

The first cycle process includes:

The target client sends the updated target clustering model (i.e., candidate model) to the server.

The server updates the parameters in the received candidate model and sends the updated candidate model to the target client.

The target client uses the first descent method to update the updated candidate model sent by the received server until the candidate model converges, and uses the converged candidate model as the personalized model of the target client.

In the model update method based on federated learning provided by the present invention, the target client participating in collaborative training uploads the data distribution information to the server. The server clusters the target client based on the data distribution information uploaded by it, and obtains the target client's membership. The cluster saves the computing resources and storage resources of the target client, reduces the communication overhead between the target client and the server, and learns an optimal personalized model (i.e., a converged candidate model) for each client.

Further, in one embodiment, updating the cluster to which the target client belongs based on the synthetic data of the target client may specifically include:

Initialize the cluster to which the target client belongs;

According to the empirical loss function of the clustering model in each cluster, obtain the empirical loss value of the clustering model in each cluster for the synthetic data of each client;

According to the empirical loss value of the clustering model of the synthetic data of each client in each cluster, the cluster to which the target client belongs is updated to determine when the K clusters reach a stable distribution state. The clients included in each cluster are the clients among the target clients that satisfy the experience loss value of the clustering model of the synthetic data in each cluster reaching the minimum value.

Further, in one embodiment, updating the cluster to which the target client belongs based on the empirical loss value of the clustering model of each client's synthetic data in each cluster may specifically include:

Execute the following second loop process until the K clusters reach a stable distribution state;

The second cycle process includes:

Update the clustering model within each cluster based on the gradient descent method;

Based on the synthetic data of the clients included in each cluster, update the empirical loss value of the updated clustering model of the clients included in each cluster within the cluster;

The cluster to which the target client belongs is updated according to the empirical loss value of the updated clustering model of the clients contained in each cluster in the cluster.

Optionally, Figure 3 is a schematic flow chart of the server provided by the present invention updating the cluster to which the target client belongs. As shown in Figure 3, the server assigns a cluster identity to each client in the target client, and the cluster identity is used to Characterizes the cluster to which the client belongs. Assume that the cluster identity of the i-th client C _i is

Initialize K cluster models on the server side, mark the Kth clustering model as θ _K , mark the number of layers of the clustering model as |θ _K |, and mark the weight parameters of the Kth clustering model as where p represents the p-th layer from the bottom of the clustering model; initialize the K-th clustering model and mark it as/> Denote the empirical loss function of the clustering model as F _K (θ _K );

Determine the empirical loss function F _K (θ _K ) of the clustering model: F _K (θ _K ) = argmin{F _K (θ _K1 ), F _k (θ _K2 ),...F _K (θ _Kn )}

The server is the empirical loss value of the synthetic data G _i of the client C _i in K clustering models: F(θ ₁ ), F(θ ₂ )...F(θ _K )

Find the minimum loss of synthetic data G _i among K models to update the cluster identity of each client Right now Until K clusters reach a stable distribution state.

When the server determines that the K clusters have reached a stable distribution state, it determines the clients included in each cluster. The clients included in each cluster are the clustering models of the target clients that satisfy the synthetic data in each cluster. The client whose experience loss value reaches the minimum value (that is, the minimum loss).

Optionally, after confirming the cluster identity in each iteration, within each cluster, the empirical loss value of the Kth clustering model is updated based on the synthetic data of the clients in the cluster.

Specifically, the server executes the following second loop process until K clusters reach a stable distribution state:

Assume that after the cluster to which the target client belongs is updated through the above process, the m clients included in the Kth cluster are C ₁ C ₂ ...C _m respectively.

The server uses gradient descent to update the clustering model in the cluster.

Within the Kth cluster, update the empirical loss of the updated clustering model within the Kth cluster based on the synthetic data G ₁ G ₂ ... _G i of the clients C ₁ C ₂ ...C _m within the cluster. value F(θ _k ), where, γ is the learning rate of clustering model weight training.

The server updates the cluster to which the target client belongs based on the empirical loss value of the updated clustering model of the clients included in each cluster until the K clusters reach a stable distribution state.

In the model update method based on federated learning provided by the present invention, the clustering work of the client is completed by the server. The target client participating in the collaborative training uploads the data distribution information to the server. The server utilizes minimization according to the uploaded data distribution information. The lossy clustering strategy divides the client. Compared with other iterative clustering federation technologies, the present invention has lower communication costs, reduces the computing resources and storage resources of the client, and at the same time utilizes the clustering provided by the present invention. This method can correctly divide the target clients, especially when the data of clients participating in federated training presents obviously heterogeneous settings, the clients will be divided into multiple different clusters.

Further, in one embodiment, receiving the candidate model sent by the target client and updating the candidate model may specifically include:

The optimization objective function of the candidate model is determined according to the sum of the first optimization objective function and the second optimization objective function. The first optimization objective function is the sum of the loss functions of the candidate models of the target client. The second optimization objective function is the sum of the loss functions of the candidate models of the target client. The optimization objective function is the sum of nonlinear functions of differences between the target clustering model received by any client and the target clustering model received by the remaining clients in the target client;

Optimize the first optimization objective function based on the gradient descent method;

Optimize the second optimization objective function based on the approximate point method;

The candidate model is updated according to the candidate model when the optimization objective function takes the minimum value.

Optionally, in each cluster that reaches a stable distribution state, obtain the clustering model of each cluster, and combine different clustering models into Distributed to each corresponding client, specifically:

When the clustering reaches a stable state, the clients C ₁ C ₂ ...C _i ...C _n are divided into K clusters, and the Kth clustering model θ _k is delivered to the clients belonging to the cluster.

Step 11: Client _Ci uses gradient descent to update the target clustering model received from the server, and sends the updated target clustering model (i.e. candidate model) to the server. Specifically:

Step 111: The client C ₁ C ₂ ...C _i ...C _n receives the target clustering model corresponding to the client's cluster identity issued by the server;

Step 112: Mark the target clustering model received by the i-th client C _i as w _i ; mark the number of layers of the target clustering model as |w _i |, and mark the weight of the target clustering model as |w i | Where p means that only local aggregation is applied on the penultimate p-th layer of the target clustering model; the loss function of the target clustering model is recorded as F( _wi );

Step 113, determine the objective function F( _wi ) of the target clustering model: F( _wi )=argmin{F(w ₁ ), F(w ₂ ),...F(w _n )}

Step 114: Determine the update method of the target clustering model, use the received target clustering model to overwrite the old local model, and then combine the local private data to update the target clustering model using gradient descent: Among them, eta is the learning rate of the local update model,/> The candidate model obtained after updating the target clustering model for client C _i ;

Step 115: Upload the updated target clustering model (ie, candidate model) to the server through the network.

Step 21: The server uses the message passing mechanism within the cluster to generate a personalized model U _i for each client based on the received candidate models sent by clients belonging to different clusters, and sends it to the corresponding client C _i ; specifically land,

Step 211, the server receives candidate models sent by clients C ₁ C ₂ ...C _i ...C _n belonging to different clusters;

Step 212, determine the personalized cloud model of the client, mark the personalized cloud model on the i-th client C _i as U _i ; mark the number of layers of the personalized cloud model as |U _i |, mark the personalized cloud model The weight of is recorded as Where p means that only local aggregation is applied on the penultimate p layer of the local model;/> Represents the parameters of the personalized cloud model belonging to the i-th client in the K-th cluster;

Step 213, determine the overall optimization objective function:

Among them, the first half of the formula represents the first optimization objective function, which is the sum of the loss functions of the target clustering models of all clients in the cluster. The second half of the formula represents the second optimization objective function, which is the attention mechanism part/> where A(||w _i -w _i || ² ) is a nonlinearity measuring the parameter difference between the target clustering model w _i received by _client C _i and the target clustering model w j received by the remaining clients C _j A function that satisfies the conditions of increasing from 0 and being differentiable.

Based on the above analysis, the optimization objective function is simplified as:

Step 214: Determine the optimization method based on the above simplified formula. Divide the optimization step into two parts: first is optimization via gradient descent Get the intermediate value U ^k and then further optimize it through the approximate point step/> According to the candidate model when the optimization objective function takes the minimum value, the candidate model is updated to obtain the updated candidate model, which is the personalized model;

Step 215: The server calculates the parameters of the personalized cloud model of each client in the cluster. Yes/> The corresponding linear combination of parameter sets is as follows:

in, is the parameter set of the target clustering model uploaded by client C _i , which is calculated by using the client's local private data, ξ _{i, 1} ,..., ξ _{i, m} corresponding to/> the weight of. (ξ _{i, 1} +...+ξ _{i, m} = 1). ξ _i,j represents the contribution weight of client C _j in the cluster to the personalized model of client C _i . /> and/> The higher the similarity between them, the greater their contribution to each other.

Step 215: Send the generated personalized cloud model to each client C ₁ C ₂ ...C _i , ..C _n .

Step 31: Repeat steps 11 and 21 to perform T rounds of training until each client's candidate model (i.e., personalized model) converges. T can be flexibly set according to the actual situation, as shown in Figure 4.

The model update method based on federated learning provided by the present invention makes full use of the similarity between clients for similar clients in the same cluster, and collaboratively trains personalized models belonging to each client for the clients. Compared with other The clustering personalized federated learning method can achieve better model performance.

Figure 5 is the second schematic flow chart of the model update method based on federated learning provided by the present invention. As shown in Figure 5, it includes:

Step 210: Send the data distribution information of the target client to the server participating in the federated learning, so that the server updates the cluster to which the target client belongs based on the synthetic data of the target client. The client includes N clients, and the synthetic data is obtained according to the received data distribution information of the target client;

Step 220: When the server determines that the initialized K clusters have reached a stable distribution state, receive the target clustering model in each cluster sent by the server, and the stable distribution state is the cluster to which the target client belongs. No longer changes, the target clustering model is determined based on the clustering model in each cluster when the initialized K clusters reach a stable distribution state;

Step 230: Execute the following first loop process until the candidate model converges. The candidate model is obtained after the target client updates the received target clustering model in each cluster;

The first cycle process includes:

Send the candidate model to the server so that the server updates the received candidate model;

Receive the updated candidate model sent by the server, and update the updated candidate model.

Optionally, it should be noted that the execution subject of the above method may be a computer device, or may be applied to a target client participating in federated learning.

Optionally, the target client may specifically be N clients randomly selected by the server participating in federated learning from the clients participating in federated learning, where N is a positive integer.

The synthetic data of the N clients can be obtained specifically from the data distribution information of the N clients. The data distribution information is the generator in each client. The generator can be specifically obtained by initializing the data deployed on the target client. Obtained after training the local model (adversarial generative network).

N clients send their corresponding data release information to the server participating in federated learning, and clustering is performed on the server side.

The server generates synthetic data corresponding to each client based on the received data distribution information uploaded by different clients (i.e., the trained generator). For example, the server inputs noise data into the trained generator to generate synthetic data for each client.

The server receives the data distribution information G 1 G 2 ...G i _of the client C ₁ C ₂ ...C _i ...C _n numbered _1, 2,...,i..., _N . ..G _n , based on the data distribution information G ₁ G ₂ ...G _i ...G _n , generate synthetic data corresponding to each client, and record the synthetic data of the i-th client information on the server as D′ _i ; Denote the synthetic data set size as |D′ _i |.

The server updates the cluster to which the target client belongs based on the synthetic data corresponding to the N clients.

The cluster can specifically be any cluster among the K clusters initialized by the server. Each cluster corresponds to a clustering model. The parameters of the clustering model are: Among them,/> Represents the parameters of the jth clustering model, K is a positive integer.

The server performs multiple iterations on the server side based on the synthetic data corresponding to the target client, and estimates the cluster to which each client belongs based on the minimum loss calculated by the empirical loss function of the clustering model in the K clusters until the initialized K clusters A stable distribution state is reached. The stable distribution state specifically means that the cluster to which each client in the target client belongs will no longer change, that is, the clients included in each cluster will no longer change.

When the server determines that the initialized K clusters have reached a stable distribution state, N clients receive the target clustering model sent by the server. The target clustering model is based on the initialization of K clusters and each cluster reaches a stable distribution state. The clustering model within the cluster is determined.

After the target client receives the target clustering model sent by the server, it updates it using the gradient descent method.

The following first loop process is executed until the candidate model converges.

The first cycle process includes:

The target client sends the updated target clustering model (i.e., candidate model) to the server.

The server updates the parameters in the received candidate model and sends the updated candidate model to the target client.

The target client uses the first descent method to update the updated candidate model sent by the received server until the candidate model converges, and uses the converged candidate model as the personalized model of the target client.

In the model update method based on federated learning provided by the present invention, the target client participating in collaborative training uploads the data distribution information to the server. The server clusters the target client based on the data distribution information uploaded by it, and obtains the target client's membership. The cluster saves the computing resources and storage resources of the target client, reduces the communication overhead between the target client and the server, and learns an optimal personalized model (i.e., a converged candidate model) for each client.

Further, in one embodiment, the method of obtaining the data distribution information of the target client may specifically include:

Based on knowledge distillation technology, pre-train teacher network;

Use the trained teacher network as the discriminator of a generative adversarial network, which is deployed on the target client;

Train the generator in the generative adversarial network based on the discriminator until the value of the loss function of the generative adversarial network is less than a preset value;

According to the trained generator, the data distribution information of the target client is determined.

Optionally, the target client trains a generative adversarial network, learns local data distribution information, and uploads the learned data distribution information to the server. Specifically:

Number the N clients participating in federated learning as C ₁ C ₂ ...C _i ...C _n ; record the data set on the i-th client as D _i ; record the number of samples included in the data set as |D _i |; mark the local generation model (i.e. generator) on the i-th client as G _i ; mark the number of layers of the local generation model as |G i |, mark the local discriminant model (i.e. generator) on the i-th client as |G _i | i.e. discriminator) marked as Denote the number of layers of the local discriminant model as/> Record the weight of the locally generated model as/> where p indicates that only local aggregation is applied on the penultimate p layer of the local model (generative adversarial network); the initialized local generative model on the i-th client is recorded as/> Mark the loss function of the generative adversarial network as/>

Introduce knowledge distillation technology and pre-train the teacher network. Mark the teacher network on the i-th client as _Ti , record the number of layers of the teacher network as |T _i |, and record the weight of the teacher network as It is trained on the teacher network based on real data, and the trained teacher network is fixed as the discriminator of the generative adversarial network/> That is, the discriminator in the generative adversarial network is initialized to the trained teacher model _Ti , that is,/>

Determine the objective function of the generative adversarial network

in, is the cross entropy loss function, y _T is the probability that the generated image belongs to each category obtained by the teacher network, t is the category with the highest probability, and is represented by a one-hot vector, cross entropy/> To measure/> Similarity to t ⁱ . The features extracted by the teacher network are expressed as/> Corresponding to the output before the fully connected layer, the L1 norm is used to measure the number of activated neurons. /> Indicates the amount of information possessed by p, where k represents the coefficient, which can be set flexibly, p′ represents a given probability vector, p′={p′ ₁ , p′ ₂ ,..., p′ _o }, and o represents The number of probabilities included in probability vector p′.

According to the objective function And a fixed discriminator/> Train the generator G _i , that is, obtain the data distribution information based on the trained generator.

Client _Ci uploads the learned data distribution information to the server.

The server receives the data distribution information G _i uploaded by the client, performs clustering on the server side, and generates synthetic data corresponding to each client based on the data distribution uploaded by different clients. The schematic diagram of the data distribution information generated through the above process is as follows: 6 shown.

Figure 7 is the third flow diagram of the federated learning model update method provided by the present invention. As shown in Figure 7, it includes:

Step 1: The target client trains a generative adversarial network, learns local data distribution information, and uploads the learned data distribution information to the server;

Step 2: The server receives the data distribution information uploaded by the client, performs clustering on the server side, and generates synthetic data corresponding to each client based on the data distribution information uploaded by different clients;

Step 3: The server initializes K clusters

Step 4: Perform multiple iterations on the server side based on the synthetic data of the target client. The cluster identity of each client is estimated based on the minimum loss calculated by the empirical loss function among the K cluster models.

Step 5: After confirming the cluster identity in each iteration, in each cluster, update the parameters of the Kth clustering model based on the synthetic data of the clients in the cluster.

Step 6: Repeat steps 4 and 5 until K clusters reach a stable distribution state;

Step 7: In each cluster that reaches a stable distribution state, obtain the clustering model of each cluster, that is, the target clustering model. According to the cluster to which the client belongs, different target clustering models are delivered to each corresponding client;

Step 8: The client uses gradient descent to update the target clustering model received from the server, and sends the updated target clustering model (i.e., candidate model) to the server;

Step 9: Based on the received candidate models of clients belonging to different clusters, the server uses the message passing mechanism to generate a personalized model for each client within the cluster, and sends it to the corresponding client;

Step 10: Repeat steps 8 and 9 to perform T rounds of training until the personalized model training of each client converges.

The model update method based on federated learning provided by the present invention learns the data distribution of the local client through the generative adversarial network and knowledge distillation technology, can accurately learn the local data distribution information, accelerate the learning and convergence of the generated model, and save the resources of the client.

The model update system based on federated learning provided by the present invention is described below. The model update system based on federated learning described below and the model update method based on federated learning described above can correspond to each other.

Figure 8 is one of the structural schematic diagrams of the model update system based on federated learning provided by the present invention. As shown in Figure 8, it is applied to servers participating in federated learning, including:

The first sending module 810 and the first update module 811;

The first sending module 810 is used to update the cluster to which the target client belongs based on the synthetic data of the target client participating in the federated learning until the initialized K clusters reach a stable distribution state, and assign each cluster to The target clustering model within is sent to the target client. The synthetic data is obtained based on the received data distribution information of the target client. The stable distribution state means that the cluster to which the target client belongs does not change. , the target client includes N clients, and the target clustering model is determined based on the clustering model in each cluster when the initialized K clusters reach a stable distribution state;

The first update module 811 is configured to perform the following first loop process until the candidate model in the target client converges. The candidate model is the target cluster in each cluster received by the target client. Obtained after updating the class model;

The first cycle process includes:

Receive the candidate model sent by the target client and update the candidate model;

The second sending module is used to send the updated candidate model to the target client, so that the target client updates the candidate model.

In the federated learning-based model update system provided by the present invention, target clients participating in collaborative training upload data distribution information to the server. The server clusters the target clients based on the uploaded data distribution information to obtain the target client's location. The cluster saves the computing resources and storage resources of the target client, reduces the communication overhead between the target client and the server, and learns an optimal personalized model (i.e., a converged candidate model) for each client.

Figure 9 is the second structural schematic diagram of the model update system based on federated learning provided by the present invention. As shown in Figure 9, it is applied to target clients participating in federated learning, including:

the second sending module 910, the receiving module 911 and the second update module 912;

The second sending module 910 is used to send the data distribution information of the target client to the server participating in the federated learning, so that the server can analyze the data distribution information of the target client according to the synthetic data of the target client. The cluster is updated, the target client includes N clients, and the synthetic data is obtained according to the received data distribution information of the target client;

The receiving module 911 is configured to receive the target clustering model in each cluster sent by the server when the server determines that the initialized K clusters have reached a stable distribution state, and the stable distribution state is the The cluster to which the target client belongs no longer changes, and the target clustering model is determined based on the clustering model in each cluster when the initialized K clusters reach a stable distribution state;

The second update module 912 is used to perform the following first loop process until the candidate model converges. The candidate model is obtained after the target client updates the received target clustering model in each cluster. ;

The first cycle process includes:

Send the candidate model to the server so that the server updates the received candidate model;

The second receiving module is configured to receive the updated candidate model sent by the server, and update the updated candidate model.

In the federated learning-based model update system provided by the present invention, target clients participating in collaborative training upload data distribution information to the server. The server clusters the target clients based on the uploaded data distribution information to obtain the target client's location. The cluster saves the computing resources and storage resources of the target client, reduces the communication overhead between the target client and the server, and learns an optimal personalized model (i.e., a converged candidate model) for each client.

Figure 10 is a schematic diagram of the physical structure of an electronic device provided by the present invention. As shown in Figure 10, the electronic device may include: a processor (processor) 1010, a communication interface (communication interface) 1011, a memory (memory) 1012 and a bus. (bus) 513, in which the processor 510, the communication interface 1011, and the memory 1012 complete communication with each other through the bus 1013. Processor 1010 can call logic instructions in memory 1012 to perform the following methods:

According to the synthetic data of the target clients participating in the federated learning, the cluster to which the target client belongs is updated until the initialized K clusters reach a stable distribution state, and the target clustering model in each cluster is sent to the Target client, the synthetic data is obtained based on the received data distribution information of the target client, the stable distribution state means that the cluster to which the target client belongs does not change, and the target client includes N clients end, the target clustering model is determined based on the clustering model in each cluster when the initialized K clusters reach a stable distribution state;

Execute the following first loop process until the candidate model in the target client converges, the candidate model is obtained after the target client updates the received target clustering model in each cluster;

The first cycle process includes:

Receive the candidate model sent by the target client and update the candidate model;

The updated candidate model is sent to the target client, so that the target client updates the candidate model.

or,

Send the data distribution information of the target client to the server participating in the federated learning, so that the server updates the cluster to which the target client belongs based on the synthetic data of the target client, where the target client includes N clients, the synthetic data is obtained according to the received data distribution information of the target client;

When the server determines that the initialized K clusters have reached a stable distribution state, receive the target clustering model in each cluster sent by the server, and the stable distribution state means that the cluster to which the target client belongs does not change. , the target clustering model is determined based on the clustering model in each cluster when the initialized K clusters reach a stable distribution state;

Execute the following first loop process until the candidate model converges, the candidate model is obtained after the target client updates the received target clustering model in each cluster;

The first cycle process includes:

Send the candidate model to the server so that the server updates the received candidate model;

Receive the updated candidate model sent by the server, and update the updated candidate model.

In addition, the above-mentioned logical instructions in the memory can be implemented in the form of software functional units and can be stored in a computer-readable storage medium when sold or used as an independent product. Based on this understanding, the technical solution of the present invention essentially or the part that contributes to the existing technology or the part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium, including Several instructions are used to cause a computer power panel (which can be a personal computer, a server, or a network power panel, etc.) to execute all or part of the steps of the methods described in various embodiments of the present invention. The aforementioned storage media include: U disk, mobile hard disk, read-only memory (ROM, Read-only Memory), random access memory (RAM, Random Access Memory), magnetic disk or optical disk and other media that can store program code. .

Further, the present invention discloses a computer program product. The computer program product includes a computer program stored on a non-transitory computer-readable storage medium. The computer program includes program instructions. When the program instructions are executed by a computer , the computer can execute the federated learning-based model update method provided by each of the above method embodiments, including, for example:

According to the synthetic data of the target clients participating in the federated learning, the cluster to which the target client belongs is updated until the initialized K clusters reach a stable distribution state, and the target clustering model in each cluster is sent to the Target client, the synthetic data is obtained based on the received data distribution information of the target client, the stable distribution state means that the cluster to which the target client belongs does not change, and the target client includes N clients end, the target clustering model is determined based on the clustering model in each cluster when the initialized K clusters reach a stable distribution state;

Execute the following first loop process until the candidate model in the target client converges, the candidate model is obtained after the target client updates the received target clustering model in each cluster;

The first cycle process includes:

Receive the candidate model sent by the target client and update the candidate model;

The updated candidate model is sent to the target client, so that the target client updates the candidate model.

or,

Send the data distribution information of the target client to the server participating in the federated learning, so that the server updates the cluster to which the target client belongs based on the synthetic data of the target client, where the target client includes N clients, the synthetic data is obtained according to the received data distribution information of the target client;

When the server determines that the initialized K clusters have reached a stable distribution state, receive the target clustering model in each cluster sent by the server, and the stable distribution state means that the cluster to which the target client belongs does not change. , the target clustering model is determined based on the clustering model in each cluster when the initialized K clusters reach a stable distribution state;

Execute the following first loop process until the candidate model converges, the candidate model is obtained after the target client updates the received target clustering model in each cluster;

The first cycle process includes:

Send the candidate model to the server so that the server updates the received candidate model;

Receive the updated candidate model sent by the server, and update the updated candidate model.

On the other hand, the present invention also provides a non-transitory computer-readable storage medium on which a computer program is stored. The computer program is implemented when executed by a processor to perform the federated learning-based model update method provided by the above embodiments. , including for example:

According to the synthetic data of the target clients participating in the federated learning, the cluster to which the target client belongs is updated until the initialized K clusters reach a stable distribution state, and the target clustering model in each cluster is sent to the Target client, the synthetic data is obtained based on the received data distribution information of the target client, the stable distribution state means that the cluster to which the target client belongs does not change, and the target client includes N clients end, the target clustering model is determined based on the clustering model in each cluster when the initialized K clusters reach a stable distribution state;

Execute the following first loop process until the candidate model in the target client converges, the candidate model is obtained after the target client updates the received target clustering model in each cluster;

The first cycle process includes:

Receive the candidate model sent by the target client and update the candidate model;

The updated candidate model is sent to the target client, so that the target client updates the candidate model.

or,

Send the data distribution information of the target client to the server participating in the federated learning, so that the server updates the cluster to which the target client belongs based on the synthetic data of the target client, where the target client includes N clients, the synthetic data is obtained according to the received data distribution information of the target client;

When the server determines that the initialized K clusters have reached a stable distribution state, receive the target clustering model in each cluster sent by the server, and the stable distribution state means that the cluster to which the target client belongs does not change. , the target clustering model is determined based on the clustering model in each cluster when the initialized K clusters reach a stable distribution state;

Execute the following first loop process until the candidate model converges, the candidate model is obtained after the target client updates the received target clustering model in each cluster;

The first cycle process includes:

Send the candidate model to the server so that the server updates the received candidate model;

Receive the updated candidate model sent by the server, and update the updated candidate model.

The system embodiments described above are only illustrative. The units described as separate components may or may not be physically separated. The components shown as units may or may not be physical units, that is, they may be located in One location, or it can be distributed across multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment. Persons of ordinary skill in the art can understand and implement the method without any creative effort.

Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by software plus a necessary general hardware platform, and of course, it can also be implemented by hardware. Based on this understanding, the part of the above technical solution that essentially contributes to the existing technology can be embodied in the form of a software product. The computer software product can be stored in a computer-readable storage medium, such as ROM/RAM, magnetic disk, optical disk, etc., including a number of instructions to cause a computer power panel (which can be a personal computer, a server, or a network power panel, etc.) to execute the methods described in various embodiments or certain parts of the embodiments.

Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention, but not to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that it can still be used Modifications are made to the technical solutions described in the foregoing embodiments, or equivalent substitutions are made to some of the technical features; however, these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims (10)

1. A model update method based on federated learning, applied to servers participating in federated learning, which is characterized by: According to the synthetic data of the target clients participating in the federated learning, the cluster to which the target client belongs is updated until the initialized K clusters reach a stable distribution state, and the target clustering model in each cluster is sent to the Target client, the synthetic data is obtained based on the received data distribution information of the target client, the stable distribution state means that the cluster to which the target client belongs does not change, and the target client includes N clients end, the target clustering model is determined based on the clustering model in each cluster when the initialized K clusters reach a stable distribution state; Execute the following first loop process until the candidate model in the target client converges, the candidate model is obtained after the target client updates the received target clustering model in each cluster; The first cycle process includes: Receive the candidate model sent by the target client and update the candidate model; The updated candidate model is sent to the target client, so that the target client updates the candidate model. 2. The model update method based on federated learning according to claim 1, applied to servers participating in federated learning, characterized in that the cluster to which the target client belongs is updated based on the synthetic data of the target client. ,include: Initialize the cluster to which the target client belongs; According to the empirical loss function of the clustering model in each cluster, obtain the empirical loss value of the clustering model in each cluster for the synthetic data of each client; According to the empirical loss value of the clustering model of the synthetic data of each client in each cluster, the cluster to which the target client belongs is updated to determine when the K clusters reach a stable distribution state. The clients included in each cluster are the clients among the target clients that satisfy the experience loss value of the clustering model of the synthetic data in each cluster reaching the minimum value. 3. The model update method based on federated learning according to claim 2, applied to servers participating in federated learning, characterized in that the clustering model in each cluster according to the synthetic data of each client The experience loss value is updated for the cluster to which the target client belongs, including: Execute the following second loop process until the K clusters reach a stable distribution state; The second cycle process includes: Update the clustering model within each cluster based on the gradient descent method; Based on the synthetic data of the clients included in each cluster, update the empirical loss value of the updated clustering model of the clients included in each cluster within the cluster; The cluster to which the target client belongs is updated according to the empirical loss value of the updated clustering model of the clients contained in each cluster in the cluster. 4. The model update method based on federated learning according to claim 1, applied to servers participating in federated learning, characterized in that: receiving the candidate model sent by the target client and updating the candidate model ,include: The optimization objective function of the candidate model is determined according to the sum of the first optimization objective function and the second optimization objective function. The first optimization objective function is the sum of the loss functions of the candidate models of the target client. The second optimization objective function is the sum of the loss functions of the candidate models of the target client. The optimization objective function is the sum of nonlinear functions of differences between the target clustering model received by any client and the target clustering model received by the remaining clients in the target client; Optimize the first optimization objective function based on the gradient descent method; Optimize the second optimization objective function based on the approximate point method; The candidate model is updated according to the candidate model when the optimization objective function takes the minimum value. 5. A model update method based on federated learning, applied to target clients participating in federated learning, characterized by: Send the data distribution information of the target client to the server participating in the federated learning, so that the server updates the cluster to which the target client belongs based on the synthetic data of the target client, where the target client includes N clients, the synthetic data is obtained according to the received data distribution information of the target client; When the server determines that the initialized K clusters have reached a stable distribution state, receive the target clustering model in each cluster sent by the server, and the stable distribution state means that the cluster to which the target client belongs does not change. , the target clustering model is determined based on the clustering model in each cluster when the initialized K clusters reach a stable distribution state; Execute the following first loop process until the candidate model converges, the candidate model is obtained after the target client updates the received target clustering model in each cluster; The first cycle process includes: Send the candidate model to the server so that the server updates the received candidate model; Receive the updated candidate model sent by the server, and update the updated candidate model. 6. The model update method based on federated learning according to claim 5, applied to target clients participating in federated learning, characterized in that the method of obtaining data distribution information of the target client includes: Based on knowledge distillation technology, pre-train teacher network; Use the trained teacher network as the discriminator of a generative adversarial network, which is deployed on the target client; Train the generator in the generative adversarial network based on the discriminator until the value of the loss function of the generative adversarial network is less than a preset value; According to the trained generator, the data distribution information of the target client is determined. 7. A model update system based on federated learning, applied to servers participating in federated learning, characterized by including: a first sending module and a first update module; The first sending module is used to update the cluster to which the target client belongs based on the synthetic data of the target client participating in the federated learning until the initialized K clusters reach a stable distribution state, and update the clusters within each cluster. The target clustering model is sent to the target client, the synthetic data is obtained according to the received data distribution information of the target client, and the stable distribution state means that the cluster to which the target client belongs does not change, The target clients include N clients, and the target clustering model is determined based on the clustering model in each cluster when the initialized K clusters reach a stable distribution state; The first update module is configured to perform the following first loop process until the candidate model in the target client converges. The candidate model is the target cluster within each cluster received by the target client. Obtained after the model is updated; The first cycle process includes: Receive the candidate model sent by the target client and update the candidate model; The second sending module is used to send the updated candidate model to the target client, so that the target client updates the candidate model. 8. A model update system based on federated learning, applied to target clients participating in federated learning, characterized by including: a second sending module, a receiving module and a second update module; The second sending module is used to send the data distribution information of the target client to the server participating in the federated learning, so that the server can analyze the cluster to which the target client belongs based on the synthetic data of the target client. Update is performed, the target client includes N clients, and the synthetic data is obtained according to the received data distribution information of the target client; The receiving module is configured to receive the target clustering model in each cluster sent by the server when the server determines that the initialized K clusters have reached a stable distribution state, and the stable distribution state is the target The cluster to which the client belongs no longer changes, and the target clustering model is determined based on the clustering model in each cluster when the initialized K clusters reach a stable distribution state; The second update module is used to perform the following first loop process until the candidate model converges. The candidate model is obtained after the target client updates the received target clustering model in each cluster; The first cycle process includes: Send the candidate model to the server so that the server updates the received candidate model; The second receiving module is configured to receive the updated candidate model sent by the server, and update the updated candidate model. 9. An electronic device, comprising a processor and a memory storing a computer program, characterized in that when the processor executes the computer program, the federated learning-based method of any one of claims 1-4 or 5-6 is implemented. model update method. 10. A non-transitory computer-readable storage medium with a computer program stored thereon, characterized in that when the computer program is executed by a processor, the federation-based storage medium as described in any one of claims 1-4 or 5-6 is implemented. Learned model update methods.