Personalized federal learning method based on client target enhancement

Abstract

The invention discloses a personalized federated learning method based on client target enhancement, and relates to the technical field of wireless communication networks. In a federated learning process, a client not only trains a model on a local data set, but also trains on target set data to obtain a target set model; the method focuses on the personalized target of each client, improves the performance of the model of each client on a specific task, calculates the personalized aggregation weight of each client according to the Euclidean distance, enables the clients with similar data distribution to have a larger weight in aggregation, can more remarkably affect the model updating, and improves the user experience. In the aspects of communication efficiency and convergence speed, only updated global model parameters need to be uploaded, communication overhead and local calculation burden are reduced, contribution of a client can be more effectively utilized by using personalized aggregation weight, and convergence speed of federal learning is accelerated.

Description

Technical Field

[0001] The present invention relates to the technical field of wireless communication networks, and in particular to a personalized federated learning method based on client target enhancement. Background Art

[0002] The proliferation of edge devices in modern society, such as mobile phones and wearable devices, has led to a rapid growth in private data from distributed sources. While the abundance of this data provides huge opportunities for machine learning applications, with the introduction of legislation such as the General Data Protection Regulation, society is increasingly concerned about data privacy, which has contributed to the growing popularity of federated learning.

[0003] Federated learning can develop a joint machine learning model on data silos in a collaborative and privacy-preserving manner. However, current federated learning has the problem of data heterogeneity in training sets between clients, which is specifically manifested as follows: the feature distribution of data sets between federated learning participants is inconsistent. For example, in Wifi-based gesture recognition, the environment between users is inconsistent; traditional federated learning has only one optimization goal, which is not well considered. Specifically, the clients participating in federated learning jointly maintain an optimization goal, and the model is optimized in one direction. However, many clients participate in aggregation and may have different purposes. Optimization solutions for this problem are provided in the existing technology, such as FedFomo. The FedFomo framework introduces a personalized model aggregation strategy, which takes into account the performance of the model and the specific situation of each client, and more finely controls the aggregation and distribution process of the model to optimize the effect of federated learning. In the FedFomo framework, each client is trained based on local data and uploads the trained model parameters. The server saves the model parameters of each client in local tasks and communication rounds. The client needs to download multiple models and aggregate these models according to certain weights. Each client aggregates the models based on the performance of other clients' models on the local validation set.

[0004] As mentioned above, in the traditional federated learning method, the server sends global model parameters to each client, and the client uses the received parameters to train on the local data set. After training, all clients send the trained model to the server. Each client only needs to download one global model parameter in each communication round. In order to solve the problem of data heterogeneity, methods similar to the FedFomo framework mentioned above require the client to download all client model parameters locally and transfer the aggregation method originally performed on the server to the local client alone. Although it can solve the problem of different data distribution to a certain extent, this method requires a large amount of data transmission and calculation, and the communication cost and calculation cost are higher. Especially when the number of clients is large or the amount of data is large, the communication overhead will increase significantly and affect the efficiency of federated learning. Therefore, there is an urgent need for a personalized federated learning method based on client target enhancement that can effectively reduce communication overhead and local computing burden while solving the problem of inconsistent data distribution between clients. Summary of the invention

[0005] 1. Technical issues to be solved

[0006] In view of the shortcomings of the prior art, the present invention provides a personalized federated learning method based on client target enhancement to solve the problems existing in the prior art such as inconsistent data distribution between clients, large communication overhead, and high local computing burden.

[0007] (II) Technical solution

[0008] To achieve the above objectives, the present invention is implemented through the following technical solutions. The present invention provides a personalized federated learning method based on client target enhancement, including:

[0009] Step 1. Initialization:

[0010] Prepare a federated learning system, including a central server and multiple clients connected to the central server;

[0011] The central server initializes the global model parameters and broadcasts them to all clients;

[0012] Define federated learning communication rounds;

[0013] Step 2: Client training and uploading, which is the first step in each update round:

[0014] Each client receives global model parameters and communication round information;

[0015] In the first round of updates, the client trains the global model on the local dataset and trains the target set model on the local target set;

[0016] The client uploads the updated global model parameters and target set model parameters to the central server;

[0017] Step 3: Central server aggregation and weight calculation, which is the second step in each update round:

[0018] The central server receives the updated global model parameters and target set model parameters uploaded by all clients;

[0019] For each client, calculate the Euclidean distance between its target set parameters and the updated global model parameters of other clients;

[0020] Generate a personalized aggregation weight matrix for each client based on the Euclidean distance calculation, ensuring that the client with a smaller distance has a larger weight;

[0021] Step 4: Generate personalized model parameters:

[0022] The central server aggregates and generates personalized model parameters for each client one by one according to the personalized aggregation weight generated for each client;

[0023] The central server sends the generated personalized model parameters to each client;

[0024] Step 5: Client update and iteration:

[0025] After receiving the personalized model parameters, each client uses the received parameters to continue training on the local data to obtain a trained local model;

[0026] The client repeats the process of receiving personalized model parameters sent by the central server and uploading updated model parameters for multiple iterations until the global model parameters of all clients in the federated learning system converge or reach the specified communication round.

[0027] The present invention is further configured as follows: in the initialization step:

[0028] The central server broadcasts the global model parameters initialized by it to each client, so that each client updates the global model according to its local data;

[0029] In the client training and uploading steps:

[0030] The communication round is used to determine whether the target set model needs to be trained and uploaded;

[0031] The target set is the motivation or task for the client to participate in federated learning aggregation training;

[0032] The central server receives the global model parameters and target set model parameters updated by the client in the first round of updates, and only accepts the global model parameters updated by the client in subsequent updates;

[0033] After training is completed, the client uploads the updated global model and target set model to the central server. Except for the first round, subsequent updates only require uploading the updated global model;

[0034] The present invention is further configured as follows: in the client training and uploading steps:

[0035] Assume that there are K clients in the network node, and each client has its corresponding training data set Target dataset Among them, N k represents the amount of data in the training set of the kth client, M k represents the amount of data owned by the k-th client target set;

[0036] The K clients jointly train a deep learning model, and the data sets between different clients are heterogeneous, that is, for any i≠j, there is a relationship

[0037] Then for client k, the training set loss function and target set loss function are: Among them, θ k For client k based on the training set D k The model parameters, θ t,k For client k based on the target set T k The target set model parameters are: Among them, w k is the personalized aggregate weight matrix generated by the central server for client k;

[0038] The present invention is further configured as follows: in the client training and uploading steps:

[0039] The client performs one or more gradient descent updates on the local data according to the local data it has, and the client performs one or more gradient descent updates on the target set data set it has.

[0040] The present invention is further configured as follows: in the central server aggregation and weight calculation steps:

[0041] When client k is updated in the first round, the training set model parameters uploaded by the client are θ k , the target set model parameters uploaded by the client are θ t,k , then for client k, the Euclidean distance from all client training set models to the client k target set model is calculated: Among them, i and k are any clients in the federated learning system;

[0042] The present invention is further configured as follows: in the step of generating a personalized model:

[0043] Generate personalized aggregate weights for each client at the central server. The personalized aggregate weight matrix for client k is w k : in, It is an intermediate function, ensuring that the smaller the distance, the greater the weight;

[0044] The formula for the aggregate weight of client k on the central server is:

[0045] After calculating the aggregation weight After that, the central server sends it to client k for further update;

[0046] The present invention is further configured as follows: in the step of generating a personalized model:

[0047] The personalized aggregation weight w is calculated on the central server except after the first round of updates. The aggregation weight w will not change in subsequent updates.

[0048] The present invention also provides a personalized federated learning system based on client target enhancement, comprising a central server and a plurality of clients connected to the central server through a network, and:

[0049] The broadcast module is used for the central server to send the initialized global model parameters. The global model parameters are used by each client to train according to its local data and are used to train the target set model in the first round of updates.

[0050] A first receiving module is used to receive global model parameters uploaded by all clients and updated after local training, and receive target set models uploaded by clients in the first round of updates, and aggregate models for each client according to personalized aggregation weights;

[0051] The second receiving module is used to receive the target set model parameters uploaded by all clients after the target set training, and calculate the Euclidean distance between the global parameters updated by the training set of each client and the model parameters of other target sets to generate personalized aggregation weights for the client;

[0052] The weight calculation module is used to calculate the Euclidean distance between the global parameters updated by the training set of each client and the model parameters of other target sets after the central server receives the local training set model parameters and the target set model parameters uploaded by all clients;

[0053] A sending module is used to send the personalized aggregation weights generated for each client on the central server and the aggregated global model parameters to the corresponding client;

[0054] A repeated iteration module, used for the client to repeatedly execute the process of determining the local update of the local optimal model according to the received updated global model parameters;

[0055] The present invention also provides a terminal device, which includes: a memory, a processor, and a control program of a personalized federated learning method based on client target enhancement stored in the memory and executable on the processor, wherein the control program of the personalized federated learning method based on client target enhancement implements the above-mentioned personalized federated learning method based on client target enhancement when executed by the processor;

[0056] The present invention also provides a storage medium, which is applied to a computer, and the storage medium stores a control program of a personalized federated learning method based on client target enhancement. When the control program of the personalized federated learning method based on client target enhancement is executed by the processor, the above-mentioned personalized federated learning method based on client target enhancement is implemented.

[0057] (III) Beneficial effects

[0058] The present invention provides a personalized federated learning method based on client target enhancement. It has the following beneficial effects:

[0059] First, the central server sends the initialized global model parameters to each client. After receiving the model parameters sent by the server, the client trains the global model based on its local training set data in the first round of updates, and also trains the target set model parameters based on its own target set data.

[0060] The client then uploads the trained global model parameters and target set model parameters to the central server. In the subsequent update process, the client only needs to upload the trained global model parameters;

[0061] Subsequently, in the first round of updates, after receiving the updated global model parameters and target set model parameters uploaded by the client, the central server calculates the Euclidean distance between the global parameters updated by the training set and the model parameters of other target sets for each client, and gives corresponding weights based on the idea that the closer the distance, the closer the distribution of data sets between clients, and generates a personalized weight matrix w. At the same time, there is no need to calculate the matrix w again in subsequent updates;

[0062] The central server then aggregates and generates personalized models for each client based on the matrix w, and then distributes the generated personalized models to the corresponding clients; the central server and the client repeat the above steps until the global model in all clients in the federated learning system converges or reaches the specified number of communications.

[0063] In summary, it can be seen that compared with the traditional federated learning method:

[0064] In the process of federated learning, the client not only trains the model on the local data set, but also trains the target set model on its target set data, focusing more on the personalized goals of each client, improving the performance of each client's model on specific tasks, and calculating the personalized aggregation weight of each client based on the Euclidean distance, so that clients with similar data distribution have a larger weight in the aggregation, which can more significantly affect the model update. In terms of communication efficiency and convergence speed, only the updated global model parameters need to be uploaded, reducing the communication overhead and local computing burden. By using personalized aggregation weights, the client's contribution can be more effectively utilized, accelerating the convergence speed of federated learning.

[0065] The problem of inconsistent data distribution between clients, high communication overhead and high local computing burden existing in the prior art is solved. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] Figure 1 A flowchart of a personalized federated learning method based on client target enhancement of the present invention;

[0067] Figure 2 A process diagram for generating personalized aggregation weights of the present invention;

[0068] Figure 3 Schematic diagram of the federated learning system of the present invention. DETAILED DESCRIPTION

[0069] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0070] Example 1

[0071] See also Figure 1-3 The present invention provides a personalized federated learning method based on client target enhancement, comprising the following steps:

[0072] S1. Initialization:

[0073] Prepare a federated learning system, including a central server and multiple clients connected to the central server;

[0074] The central server initializes the global model parameters and broadcasts them to all clients;

[0075] Define federated learning communication rounds;

[0076] The central server broadcasts the global model parameters initialized by it to each client, so that each client updates the global model according to its local data;

[0077] S2, client training and uploading, which is the first step in each round of update:

[0078] Each client receives global model parameters and communication round information;

[0079] In the first round of updates, the client trains the global model on the local dataset and trains the target set model on the local target set;

[0080] The client uploads the updated global model parameters and target set model parameters to the central server;

[0081] The communication round is used to determine whether the target set model needs to be trained and uploaded;

[0082] The target set is the motivation or task for the client to participate in federated learning aggregation training;

[0083] The central server receives the global model parameters and target set model parameters updated by the client in the first round of updates, and only accepts the global model parameters updated by the client in subsequent updates;

[0084] After training is completed, the client uploads the updated global model and target set model to the central server. Except for the first round, subsequent updates only require uploading the updated global model;

[0085] Assume that there are K clients in the network node, and each client has its corresponding training data set Target dataset Among them, N k represents the amount of data in the training set of the kth client, M k represents the amount of data owned by the k-th client target set;

[0086] K clients jointly train a deep learning model, and the datasets between different clients are heterogeneous, that is, for any i≠j, there is a relationship

[0087] Then for client k, the training set loss function and target set loss function are: Among them, θ k For client k based on the training set D k The model parameters, θ t,k For client k based on the target set T k The target set model parameters are: Among them, w k is the personalized aggregate weight matrix generated by the central server for client k;

[0088] The client performs one or more gradient descent updates on the local data based on the local data it has. The client performs one or more gradient descent updates on the target set data set it has.

[0089] The client needs to receive the global model parameters, i.e., the initial model parameters, broadcasted from the central server. At the same time, the client also receives the communication round information in order to make judgments and decisions in the subsequent training process and perform dual training in the first round of updates. In the first round of updates, the client not only updates the global model parameters on the local data set, but also trains on the local target set to obtain the target set model. The target set model is the driving force for the client to participate in federated learning and reflects the personalized needs of the client's tasks. The purpose of dual training is to establish a model that adapts to the client's own task goals at an early stage. After completing the dual training, the client uploads the updated global model parameters and target set model parameters to the central server. In this way, the central server can obtain the client's local training results and personalized target information, enhance personalization, and more effectively transmit information during the communication process.

[0090] S3, central server aggregation and weight calculation, which is the second step in each round of update:

[0091] The central server receives the updated global model parameters and target set model parameters uploaded by all clients;

[0092] For each client, calculate the Euclidean distance between its target set parameters and the updated global model parameters of other clients;

[0093] A personalized aggregation weight matrix is ​​generated for each client based on the Euclidean distance calculation, ensuring that the client with a smaller distance has a larger weight.

[0094] The central server needs to receive the updated global model parameters and target set model parameters uploaded by all clients for subsequent calculation and aggregation operations. For each client, the central server calculates the Euclidean distance between its target set parameters and the updated global model parameters of other clients. This Euclidean distance reflects the degree of difference between clients in the target set. The smaller the distance, the more similar the target set models of the two are. Based on the calculated Euclidean distance, the central server calculates the personalized aggregation weight matrix. The weights in this matrix represent the contribution of each client in the model aggregation process. Clients with smaller distances will obtain larger weights, which is more conducive to model aggregation and achieves more targeted model aggregation.

[0095] When client k is updated in the first round, the training set model parameters uploaded by the client are θ k , the target set model parameters uploaded by the client are θ t,k , then for client k, the Euclidean distance from all client training set models to the client k target set model is calculated: Among them, i and k are any clients in the federated learning system.

[0096] S4. Generate personalized model parameters:

[0097] The central server aggregates and generates personalized model parameters for each client one by one according to the personalized aggregation weight generated for each client;

[0098] The central server sends the generated personalized model parameters to each client;

[0099] Since the smaller the Euclidean distance, the more similar the two models are, which means that the client distributions of the two are more similar, a personalized aggregation weight is generated for each client on the central server. The personalized aggregation weight matrix for client k is w k : in, It is an intermediate function, ensuring that the smaller the distance, the greater the weight;

[0100] The formula for the aggregate weight of client k on the central server is:

[0101] After calculating the aggregation weight After that, the central server sends it to client k for further update;

[0102] The personalized aggregation weight w is calculated on the central server except after the first round of updates. The aggregation weight w will not change in subsequent updates.

[0103] The central server performs weighted averaging of the updated global model parameters of each client based on the personalized aggregation weight matrix calculated previously. This weighted average takes into account the contribution of each client in the model aggregation, thereby generating personalized model parameters for each client. At the same time, the central server sends the generated personalized model parameters to each client. Each client can use this personalized model parameter to perform local model updates and training. Through personalized model parameters, the client can better adapt to its own task objectives and data distribution, thereby improving the performance of the model on local tasks and achieving personalized model updates and training.

[0104] S5. Client updates and iterations:

[0105] After each client receives the personalized model parameters, it uses the received parameters to continue training on the local data to obtain a trained local model; that is, a better local model.

[0106] The client repeats the process of receiving personalized model parameters sent by the central server and uploading updated model parameters for multiple iterations until the global model parameters of all clients in the federated learning system converge or reach the specified communication round.

[0107] The present invention also provides a personalized federated learning system based on client target enhancement, comprising a central server and a plurality of clients connected to the central server through a network, and:

[0108] The broadcast module is used for the central server to send the initialized global model parameters. The global model parameters are used by each client to train according to its local data and are used to train the target set model in the first round of updates.

[0109] A first receiving module is used to receive global model parameters uploaded by all clients and updated after local training, and receive target set models uploaded by clients in the first round of updates, and aggregate models for each client according to personalized aggregation weights;

[0110] The second receiving module is used to receive the target set model parameters uploaded by all clients after the target set training, and calculate the Euclidean distance between the global parameters updated by the training set of each client and the model parameters of other target sets to generate personalized aggregation weights for the client;

[0111] The weight calculation module is used to calculate the Euclidean distance between the global parameters updated by the training set of each client and the model parameters of other target sets after the central server receives the local training set model parameters and the target set model parameters uploaded by all clients;

[0112] A sending module is used to send the personalized aggregation weights generated for each client on the central server and the aggregated global model parameters to the corresponding client;

[0113] The repeated iteration module is used for the client to repeatedly execute the process of determining the local update of the local optimal model according to the received updated global model parameters.

[0114] The present invention also provides a terminal device, which includes: a memory, a processor, and a control program of a personalized federated learning method based on client target enhancement, which is stored in the memory and can be run on the processor. When the control program of the personalized federated learning method based on client target enhancement is executed by the processor, the above-mentioned personalized federated learning method based on client target enhancement is implemented.

[0115] The present invention also provides a storage medium, which is applied to a computer, and stores a control program of a personalized federated learning method based on client target enhancement. When the control program of the personalized federated learning method based on client target enhancement is executed by a processor, the above-mentioned personalized federated learning method based on client target enhancement is implemented.

[0116] Example 2

[0117] The main difference from Example 1 is that in the weight calculation module, the calculation process of the personalized aggregation weight matrix w is that in the first round of updates, the central server receives the global model parameters updated by the training set and the target set model parameters uploaded by the client, and calculates the Euclidean distance between the global parameters updated by the training set of each client and the other target set model parameters, and generates a personalized aggregation weight matrix w for each client based on the idea that the smaller the distance, the more similar the client data distribution.

[0118] By measuring the Euclidean distance between the global parameters updated by the training set of each client and the model parameters of other target sets, clients with smaller distances or similar data distributions are given greater aggregation weights during the central server aggregation process to reduce the impact of data heterogeneity between clients. At the same time, from the perspective of the client, a personalized model is generated with the client's goals as the center, and federated learning is applied to more practical scenarios.

[0119] In this embodiment, by measuring the Euclidean distance between the global parameters of each client updated by the training set and the target set model parameters of other clients, the data distribution differences between the clients are quantified. According to the principle that the smaller the distance, the more similar the data distribution, a personalized aggregation weight matrix is ​​generated for each client, thereby reducing the impact caused by data heterogeneity between different clients and better utilizing cross-client information. By assigning larger aggregation weights to clients with smaller distances (i.e., more similar data distributions), the central server generates a personalized aggregation model for each client, which means that during the model update and training process, each client can pay more attention to information related to its target task and data distribution, thereby improving the performance of the model on local tasks.

[0120] This embodiment applies federated learning to more practical scenarios. Through data distribution differences, personalized aggregation weights, and more sophisticated model customization, federated learning can better adapt to the needs of different clients and tasks, thereby playing a role in a wider range of application scenarios.

[0121] In view of the above, in this application:

[0122] In the process of federated learning, the client not only trains the model on the local data set, but also trains the target set model on its target set data, focusing more on the personalized goals of each client, improving the performance of each client's model on specific tasks, and calculating the personalized aggregation weight of each client based on the Euclidean distance, so that clients with similar data distribution have greater weight in the aggregation, which can more significantly affect the model update. In terms of communication efficiency and convergence speed, only the updated global model parameters need to be uploaded, reducing the communication overhead. By using personalized aggregation weights, the client's contribution can be more effectively utilized, accelerating the convergence speed of federated learning.

[0123] In the description of the embodiments of the present invention, it will be understood by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A personalized federated learning method based on client target enhancement, characterized in that: include: Step 1. Initialization: Prepare a federated learning system, including a central server and multiple clients connected to the central server; The central server initializes the global model parameters and broadcasts them to all clients; Define federated learning communication rounds; Step 2: Client training and uploading, which is the first step in each update round: Each client receives global model parameters and communication round information; In the first round of updates, the client trains the global model on the local dataset and obtains the target set model by training on the local target set; The client uploads the updated global model parameters and target set model parameters to the central server; Step 3: Central server aggregation and weight calculation, which is the second step in each update round: The central server receives the updated global model parameters and target set model parameters uploaded by all clients; For each client, calculate the Euclidean distance between its target set parameters and the updated global model parameters of other clients; Generate a personalized aggregation weight matrix for each client based on the Euclidean distance calculation, ensuring that the client with a smaller distance has a larger weight; Step 4: Generate personalized model parameters: The central server aggregates and generates personalized model parameters for each client one by one according to the personalized aggregation weight generated for each client; The central server sends the generated personalized model parameters to each client; Step 5: Client update and iteration: After receiving the personalized model parameters, each client uses the received parameters to continue training on the local data to obtain a trained local model; The client repeats the process of receiving personalized model parameters sent by the central server and uploading updated model parameters for multiple iterations until the global model parameters of all clients in the federated learning system converge or reach the specified communication round.

2. According to claim 1, a personalized federated learning method based on client target enhancement is characterized in that: In the initialization step: The central server broadcasts the global model parameters initialized by it to each client, so that each client updates the global model according to its local data; In the client training and uploading steps: The communication round is used to determine whether the target set model needs to be trained and uploaded; The target set is the motivation or task for the client to participate in federated learning aggregation training; The central server receives the global model parameters and target set model parameters updated by the client in the first round of updates, and only accepts the global model parameters updated by the client in subsequent updates; After training is completed, the client uploads the updated global model and target set model to the central server. Except for the first round, subsequent updates only require uploading the updated global model.

3. According to claim 1, a personalized federated learning method based on client target enhancement is characterized in that: In the client training and uploading steps: Assume that there are K clients in the network node, and each client has its corresponding training data set Target dataset Among them, N k represents the amount of data in the training set of the kth client, M k represents the amount of data owned by the k-th client target set; The K clients jointly train a deep learning model, and the data sets between different clients are heterogeneous, that is, for any i≠j, there is a relationship Then for client k, the training set loss function and target set loss function are: Among them, θ k For client k based on the training set D k The model parameters, θ t,k For client k based on the target set T k The target set model parameters are: Among them, w k is the personalized aggregate weight matrix generated by the central server for client k.

4. According to claim 2, a personalized federated learning method based on client target enhancement is characterized in that: In the client training and uploading steps: The client performs training based on its local data by performing one or more gradient descent updates on the local data, and the client performs training based on its target set data set by performing one or more gradient descent updates on the local data.

5. According to claim 1, a personalized federated learning method based on client target enhancement is characterized in that: In the central server aggregation and weight calculation steps: When client k is updated in the first round, the training set model parameters uploaded by the client are θ k , the target set model parameters uploaded by the client are θ t,k , then for client k, the Euclidean distance from all client training set models to the client k target set model is calculated: Among them, i and k are any clients in the federated learning system.

6. According to claim 1, a personalized federated learning method based on client target enhancement is characterized in that: In the step of generating a personalized model: Generate personalized aggregation weights for each client in the central server. The personalized aggregation weight matrix for client k is w k : in, It is an intermediate function, ensuring that the smaller the distance, the greater the weight; The formula for the aggregate weight of client k on the central server is: After calculating the aggregation weight After that, the central server sends it to client k for further updating.

7. A personalized federated learning method based on client target enhancement according to claim 6, characterized in that: In the step of generating a personalized model: The personalized aggregation weight w is calculated on the central server except after the first round of updates, and the aggregation weight w will not change in subsequent updates.

8. A personalized federated learning system based on client target enhancement, characterized in that: It includes a central server and several clients connected to the central server through a network, as well as: The broadcast module is used for the central server to send the initialized global model parameters. The global model parameters are used by each client to train according to its local data and are used to train the target set model in the first round of updates. A first receiving module is used to receive global model parameters uploaded by all clients and updated after local training, and receive target set models uploaded by clients in the first round of updates, and aggregate models for each client according to personalized aggregation weights; The second receiving module is used to receive the target set model parameters uploaded by all clients after the target set training, and calculate the Euclidean distance between the global parameters updated by the training set of each client and the model parameters of other target sets to generate personalized aggregation weights for the client; The weight calculation module is used to calculate the Euclidean distance between the global parameters updated by the training set of each client and the model parameters of other target sets after the central server receives the local training set model parameters and the target set model parameters uploaded by all clients; A sending module is used to send the personalized aggregation weights generated for each client on the central server and the aggregated global model parameters to the corresponding client; The repeated iteration module is used for the client to repeatedly execute the process of determining the local update of the local optimal model according to the received updated global model parameters.

9. A terminal device, characterized in that: The device includes: a memory, a processor, and a control program of a personalized federated learning method based on client target enhancement, which is stored in the memory and can be run on the processor. When the control program of the personalized federated learning method based on client target enhancement is executed by the processor, a personalized federated learning method based on client target enhancement as described in any one of claims 1-7 is implemented.

10. A storage medium, characterized in that: The medium is applied to a computer, and a control program of a personalized federated learning method based on client target enhancement is stored on the storage medium. When the control program of the personalized federated learning method based on client target enhancement is executed by the processor, a personalized federated learning method based on client target enhancement as described in any one of claims 1-7 is implemented.

Images