Clustering federated learning method based on model decomposition and electronic equipment

Abstract

The invention relates to a clustering federated learning method based on model decomposition and electronic equipment, and the method is realized based on model decomposition and low-rank similarity, and comprises the following steps: model decomposition and initialization: splitting a client model into a shallow embedding model and a deep decision model, initializing global embedding parameters and clustering center parameters of the server; based on a clustering algorithm of low-rank model similarity, performing low-rank decomposition on the client model splicing matrix through a dimension reduction technology to obtain a low-rank matrix, and calculating the similarity of the low-rank matrix; a knowledge integration and parameter updating mechanism: optimizing and updating client model parameters in combination with global embedded knowledge and clustering center knowledge in the step 1; and dynamically updating the parameters of the clustering distribution and clustering center model: dynamically adjusting the clustering distribution matrix based on the low-rank representation of the low-rank matrix, and updating the parameters of the global embedding and clustering center. Compared with the prior art, the method has the advantages of improving the performance and efficiency of federal learning in a heterogeneous environment and the like.

Description

Technical Field

[0001] The present invention relates to the field of distributed machine learning technology, and in particular to a clustering federated learning method and electronic device based on model decomposition. Background Art

[0002] Federated learning, a distributed machine learning paradigm that protects data privacy, holds broad application prospects in areas such as smart healthcare, the Internet of Things, and financial technology. However, in practice, traditional federated learning algorithms face significant challenges due to significant differences in client data distribution and heterogeneous device computing capabilities.

[0003] Existing technologies such as single global model methods such as FedAvg face problems such as degraded model performance and slow convergence in heterogeneous data environments; and although multi-center clustering methods (such as IFCA and FeSEM) have improved, they still have two key limitations: one is the knowledge isolation problem. Traditional clustering algorithms strictly limit knowledge sharing to within the cluster and lack an effective mechanism to integrate global knowledge with local training, resulting in limited model performance; the other is the computational complexity problem. As the number of edge devices participating in training increases, the overhead of computing the similarity between device models and clustering centers shows a significant growth trend, limiting the efficiency of large-scale deployment. There is an urgent need for an efficient clustering federated learning method that can optimize knowledge sharing strategies and reduce computational complexity.

[0004] After searching, Chinese invention patent application publication number CN117422154A discloses a method for implementing cluster federated learning with feature alignment, including: adding a global model to the training process for knowledge transfer (i.e., feature alignment), uploading the weight updates of the local models and their models to the server; the server obtains the local models of each client and their weight updates, calculates the client similarity based on the parameter updates of each client's local model, and clusters the clients using spectral clustering to obtain clustering results; the server evaluates the clustering results using the Calinski-Harabasz metric, selects the best clustering result, and weightedly aggregates the client models in each cluster to obtain k cluster models; and weightedly aggregates the k cluster models to obtain a global model; in the next round of communication, the client downloads the cluster models and the weighted aggregated global model and uses the cluster models to update the local model. This existing patent application suffers from knowledge isolation issues and high computational complexity.

[0005] How to achieve high-performance clustering federated learning with low computational complexity has become a technical problem that needs to be solved. Summary of the Invention

[0006] The purpose of the present invention is to overcome the defects of the above-mentioned prior art and provide a clustering federated learning method and electronic device based on model decomposition.

[0007] The purpose of the present invention can be achieved by the following technical solutions:

[0008] According to one aspect of the present invention, a clustering federated learning method based on model decomposition is provided. The method is based on model decomposition and low-rank similarity and includes the following steps:

[0009] Step 1: Model decomposition and initialization: Split the client model into a shallow embedding model and a deep decision model, and initialize the server's global embedding parameters and cluster center parameters;

[0010] Step 2: Based on the low-rank model similarity clustering algorithm, the client model splicing matrix is ​​decomposed into a low-rank matrix by dimensionality reduction technology, and the low-rank matrix similarity is calculated;

[0011] Step 3: Knowledge integration and parameter update mechanism: Combine the global embedding knowledge and cluster center knowledge from step 1 to optimize and update the client model parameters;

[0012] Step 4: Dynamic update of cluster assignment and cluster center model parameters: Dynamically adjust the cluster assignment matrix based on the low-rank representation of the low-rank matrix, and update the global embedding and cluster center parameters.

[0013] Preferably, in step 1, the shallow embedding model is responsible for processing the low-dimensional features of the input data and focusing on the global structure of the data; the deep decision model focuses on capturing device-specific data distribution and is suitable for personalized learning.

[0014] Preferably, the calculation process of the low-rank model similarity includes:

[0015] The server samples the client's local model and selects some clients for offline calculation;

[0016] Flatten and concatenate the local models of the selected clients into the matrix W d ;

[0017] For the matrix W d Apply dimensionality reduction technology to perform low-rank decomposition and obtain a low-rank matrix M;

[0018] Calculate the similarity between the client model and the cluster center, that is, the low-rank matrix similarity.

[0019] More preferably, the process of knowledge integration and parameter updating in step 3 includes:

[0020] Server Aggregation Client Model w i Forming a global embedding g e ;

[0021] The server calculates the kth cluster center model where r i (k) is the cluster assignment of the i-th client model at the k-th cluster center, and m is the number of client models;

[0022] The server will globally embed g e and the corresponding cluster center model Send to each client;

[0023] The client updates the local model parameters based on the received global embedding and cluster center model.

[0024] More preferably, the local model parameter update is performed based on a triple optimization objective, wherein the triple optimization objective includes a supervision loss L1, an intra-cluster regularization loss L2, and a global regularization loss L3;

[0025] Minimizing the sum of the three losses is used as the comprehensive optimization objective function for updating local model parameters.

[0026] More preferably, the supervision loss is based on the task-related loss of the client's local data to ensure the model's fit to the local data;

[0027] The intra-cluster regularization loss is used to control the closeness between the client model and the model of the cluster center to which it belongs, and to maintain the consistency number with the cluster center;

[0028] The global regularization loss is used to promote global knowledge sharing and ensure the consistency of shallow embedding models with the global embedding.

[0029] Preferably, the dynamic updating of cluster center model parameters includes:

[0030] For cluster center k, calculate the weighted average of all client models assigned to the cluster to update the cluster center model

[0031]

[0032] Among them, r i (k) For the i-th client model w i Cluster assignment at the kth cluster center, where m is the number of client models;

[0033] Based on shallow embedding model Update the global embedding g e :

[0034] The updated cluster center model and global embedding are broadcast to the corresponding clients for the next round of local training.

[0035] More preferably, the method further includes updating the local model of the client, the process of which includes:

[0036] The client receives the global embedding g sent by the server e and cluster center model

[0037] Perform multiple rounds of local training, updating the client model in each round according to the following formula:

[0038]

[0039] Among them, w i 'For the updated i-th client model, is the supervised loss gradient, μ is the intra-cluster regularization parameter, λ is the global regularization parameter, and η is the learning rate;

[0040] The updated client model w i and low-rank representation M·w i Send back to the server, where M is a low-rank matrix.

[0041] Preferably, the updating of the cluster allocation includes: calculating the similarity matrix between the client and each cluster center based on the low-rank model similarity; and updating the cluster allocation matrix according to the principle of maximum similarity.

[0042] According to another aspect of the present invention, an electronic device is provided, comprising a memory and a processor, wherein a computer program is stored in the memory, and the method described above is implemented when the processor executes the program.

[0043] Compared with the prior art, the present invention has the following beneficial effects:

[0044] 1) The present invention realizes the organic unity of global knowledge sharing and local personalization through a model decomposition strategy, solving the problem of knowledge isolation in traditional clustering federated learning; it uses low-rank similarity calculation to reduce the computational complexity from O(m×K×dim(w)) to O(m×K×D), where D is much smaller than dim(w), greatly reducing the computational overhead; this method improves the performance and efficiency of federated learning in heterogeneous environments, and is suitable for distributed learning scenarios such as edge computing and mobile devices with limited resources and heterogeneous data, and has broad application prospects.

[0045] 2) The present invention updates local model parameters through a triple optimization objective strategy (supervisory loss, intra-cluster regularization, and global regularization), ensuring that the model retains personalized features while fully utilizing global knowledge to enhance generalization capabilities, thereby significantly improving the performance of federated learning in heterogeneous environments.

[0046] 3) The low-rank similarity matrix of the present invention dynamically adjusts the clustering allocation of the client, adapts to the dynamic changes in data distribution, and improves the adaptability and convergence speed of the model to heterogeneous environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 This is the architecture diagram of the clustering federated learning system based on model decomposition;

[0048] Figure 2 This is the flow chart of the FedSFC algorithm;

[0049] Figure 3 Schematic diagram of model decomposition and knowledge integration mechanism;

[0050] Figure 4 Schematic diagram for low-rank model similarity calculation;

[0051] Figure 5 This is a comparison chart of model performance improvement in a heterogeneous data environment;

[0052] Figure 6 This is a diagram analyzing the effect of reducing computational complexity;

[0053] Figure 7 Schematic diagram of the flow of the clustering federated learning method based on model decomposition in the present invention. DETAILED DESCRIPTION

[0054] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts should fall within the scope of protection of the present invention.

[0055] Example 1

[0056] This embodiment relates to a clustering federated learning method based on model decomposition (Federated Split Feature Coordination, FedSFC). This method is based on model decomposition and low-rank similarity calculation, and is used to solve the problems of knowledge isolation and computational complexity faced by federated learning in heterogeneous data environments.

[0057] The method includes:

[0058] Model decomposition and knowledge integration mechanism: The client model is divided into two complementary components: a shallow embedding model and a deep decision model. The shallow embedding model captures the global structural characteristics of the data, while the deep decision model focuses on the device-specific data distribution, achieving an organic combination of global knowledge sharing and local personalization.

[0059] Low-rank model similarity calculation: Dimensionality reduction technology is used to perform low-rank decomposition of device model parameters, significantly reducing the computational complexity of the clustering algorithm in the stage of calculating the low-rank similarity of the model, thereby improving system scalability.

[0060] Triple Optimization Objective Strategy: Introducing three optimization objectives during the model update phase: supervised loss, intra-cluster regularization, and global regularization. This ensures that the model retains personalized features while fully leveraging global knowledge to improve generalization capabilities.

[0061] Dynamic cluster allocation mechanism: Dynamically adjust the client's cluster allocation based on the low-rank similarity matrix to adapt to the dynamic changes in data distribution and improve the model's adaptability to heterogeneous environments.

[0062] like Figure 2 , the method comprising:

[0063] Step 1: System initialization and model decomposition

[0064] 1.1 Global initialization: The server randomly initializes K cluster center model parameters And send the initial model to all clients as the client model;

[0065] 1.2 Model decomposition: Decompose the client model w i Decomposition into shallow embedding models and deep decision model h i Two components with complementary functions;

[0066] 1.3 Low-rank matrix calculation: The server samples the local model parameters from the client, constructs the matrix, and obtains the low-rank matrix M∈R through dimensionality reduction techniques (such as PCA) dim(w)×D .

[0067] Step 2: Federated Training and Knowledge Sharing

[0068] 2.1 Client selection: A portion of clients S are randomly selected in each round of training d Participate in training;

[0069] 2.2 Parameter distribution: The server sends the corresponding cluster center parameters to the selected client and global embedding g e ;

[0070] 2.3 Local Update: The client updates the model based on the triple optimization objectives, which include:

[0071] Supervision loss: Task-related loss based on client-side local data ensures the model fits the local data;

[0072] Intra-cluster regularization: Controls the closeness between the client model and the cluster center model to which it belongs, maintaining consistency with the cluster center. μ is the intra-cluster regularization parameter.

[0073] Global Regularization: Promote global knowledge sharing and ensure the consistency between the shallow embedding model and the global embedding, where g e is a global embedding, is a shallow embedding model, and λ is a global regularization parameter that controls the alignment strength between the local model and the global knowledge.

[0074] Comprehensive optimization objective function: like Figure 3 .

[0075] 2.4 Parameter upload: The client uploads the updated model parameters w i and low-rank model representation M·w i Send back to the server.

[0076] Step 3: Cluster update and knowledge integration

[0077] 3.1 Low-rank similarity calculation: Figure 4 , the server calculates the similarity between the client model and each cluster center using the following formula:

[0078]

[0079] 3.2 Cluster assignment update: Update the client's cluster assignment r according to the similarity matrix i (k) ;

[0080] 3.3 Knowledge Integration: Updating Global Embedding and the parameters of each cluster center

[0081] 3.4 Iterative optimization: Repeat steps 2 and 3 until the model converges or reaches the preset training rounds.

[0082] Step 4: Model Evaluation and Deployment

[0083] 4.1 Performance evaluation: Evaluate the model’s accuracy, F1 score, and other indicators on the test set;

[0084] 4.2 Computational Efficiency Analysis: Comparison of the differences between traditional clustering methods and the proposed method in terms of computational complexity and communication overhead;

[0085] 4.3 Model deployment: Deploy the trained model to the actual application scenario and verify the effect.

[0086] Example 2

[0087] This embodiment also relates to a clustering federated learning method based on model decomposition, such as Figure 7 , including the following steps:

[0088] S1, model decomposition and initialization, split the client model into a shallow embedding model and a deep decision model, and initialize the server's global embedding parameters and cluster center parameters;

[0089] S2, a clustering algorithm based on low-rank model similarity, which reduces computational complexity through dimensionality reduction technology and efficiently allocates clients to different cluster centers;

[0090] S3, knowledge integration and parameter update mechanism: combines global embedding knowledge and cluster center knowledge to optimize and update the client local model;

[0091] S4. Multi-level regularization loss function design: including supervision loss, intra-cluster regularization loss and global regularization loss;

[0092] S5. Dynamic update of cluster assignment and cluster center model parameters: Dynamically adjust the cluster assignment matrix based on low-rank representation, and update the global embedding and cluster center parameters.

[0093] The implementation process of model decomposition in S1 includes: i Split into shallow embedding model and deep decision model h i ,satisfy The shallow embedding model Responsible for processing the low-dimensional features of the input data and focusing on the global structure of the data; deep decision model h i Focuses on capturing device-specific data distribution, suitable for personalized learning.

[0094] The calculation process of low-rank model similarity in S2 includes:

[0095] S21. The server samples the client's local model and selects some clients for offline calculation;

[0096] S22. Flatten the local model of the selected client and splice it into a matrix Where dim(w) is the parameter dimension of the local model after the parameter vector w is flattened, S d For the selected client collection;

[0097] S23, matrix W d Apply dimensionality reduction technology to perform low-rank decomposition and obtain a low-rank matrix Where D is the dimension of the low-rank space;

[0098] S24. Calculate the similarity between the client model and the cluster center using the following formula, i.e., the low-rank matrix similarity:

[0099]

[0100] Among them, ||||2 represents the L2 norm, w i is the local model parameter of the i-th client, is the model parameter vector of the k-th cluster center, and M is a low-rank matrix.

[0101] The knowledge integration and parameter update mechanism in S3 includes the following sub-steps:

[0102] S31, Server Aggregation Client Model Forming a global embedding m is the number of local models;

[0103] S32, the server calculates the kth cluster center model where r i (k) Assignment for clustering;

[0104] S33, the server will globally embed g e and the corresponding cluster center model Send to each client;

[0105] S34. The client updates the local model parameters based on the received global embedding and cluster center model.

[0106] The update process of cluster assignment in S5 includes:

[0107] S51. Calculate the similarity matrix between the client and each cluster center based on the low-rank model similarity;

[0108] S52. Update the cluster allocation matrix according to the principle of maximum similarity:

[0109]

[0110] in, is the kth cluster center model, w i is the i-th client model.

[0111] S5 also includes a parameter update method for the cluster center model:

[0112] S53. For cluster center k, calculate the weighted average of all client models assigned to the cluster:

[0113] S54. Update global embedding parameters:

[0114] S55. Broadcast the updated parameters to the corresponding client for the next round of local training.

[0115] The process of updating the client local model in S3 includes:

[0116] The client receives the global embedding g sent by the server e and the cluster center parameters

[0117] Perform multiple rounds of local training, and update the client local model parameters according to the following formula in each round:

[0118]

[0119] The updated client local model parameter w i and low rank M·w i Send back to the server.

[0120] The computational complexity of this method is reduced from O(m×K×dim(w)) to O(m×K×D)), where m is the number of clients, K is the number of cluster centers, dim(w) is the model parameter dimension, D is the low-rank representation dimension and D< <dim(w)。

[0121] like Figure 5 and Figure 6 Experimental verification shows that the proposed method achieves an accuracy of 94.9% and 95.6% on the CIFAR-10 and MNIST datasets, respectively, which is significantly improved compared with the traditional methods FedAvg, IFCA, FedPer and FedSoft. It effectively solves the problems of knowledge sharing and computational overhead while maintaining high model accuracy, and provides a new theoretical and practical perspective for federated learning in non-independent and identically distributed environments.

[0122] Example 3

[0123] This embodiment also relates to a clustering federated learning system based on model decomposition, such as Figure 1 ,The system includes a server and a client. The server includes a model decomposition module, a low-rank similarity ,calculation module, a clustering assignment module, and a knowledge integration module.

[0124] Server-side model decomposition module: responsible for decomposing the client model into a shallow embedding model and a deep decision model;

[0125] Low-rank similarity calculation module: reduces the complexity of model similarity calculation through dimensionality reduction technology, obtains a low-rank similarity matrix, and improves system scalability;

[0126] Cluster allocation module: Dynamically adjusts the client's cluster allocation based on the low-rank similarity matrix to adapt to the dynamic changes in data distribution and improve the model's adaptability to heterogeneous environments;

[0127] Knowledge Integration Module: This module integrates global knowledge with intra-cluster knowledge to optimize the model training process. This includes: the server aggregates client local models to form a global embedding; the server calculates the cluster center model; the server sends the global embedding and the corresponding cluster center model to each client; and the client updates the local model parameters based on the received global embedding and cluster center model.

[0128] The electronic device of the present invention includes a central processing unit (CPU), which can perform various appropriate actions and processes according to computer program instructions stored in a read-only memory (ROM) or loaded from a storage unit into a random access memory (RAM). In the RAM, various programs and data required for device operation can also be stored. The CPU, ROM, and RAM are connected to each other via a bus. An input / output (I / O) interface is also connected to the bus.

[0129] Many components in a device are connected to the I / O interface, including: input units, such as a keyboard and mouse; output units, such as various types of displays and speakers; storage units, such as magnetic disks and optical disks; and communication units, such as network cards, modems, and wireless communication transceivers. The communication unit allows the device to exchange information / data with other devices via computer networks such as the Internet and / or various telecommunication networks.

[0130] The processing unit performs the various methods and processes described above. For example, in some embodiments, the method can be implemented as a computer software program, which is tangibly contained in a machine-readable medium, such as a storage unit. In some embodiments, part or all of the computer program can be loaded and / or installed on the device via a ROM and / or a communication unit. When the computer program is loaded into RAM and executed by the CPU, one or more steps of the method described above can be performed. Alternatively, in other embodiments, the CPU can be configured to execute the method in any other appropriate manner (e.g., by means of firmware).

[0131] The functions described above herein may be performed, at least in part, by one or more hardware logic components. For example, and without limitation, exemplary types of hardware logic components that may be used include: field programmable gate arrays (FPGAs), application specific integrated circuits (ASICs), application specific standard products (ASSPs), systems on chip (SOCs), complex programmable logic devices (CPLDs), and the like.

[0132] The program code for implementing the method of the present invention can be written in any combination of one or more programming languages. Such program code can be provided to a processor or controller of a general-purpose computer, a special-purpose computer, or other programmable data processing device so that when the program code is executed by the processor or controller, the functions / operations specified in the flow chart and / or block diagram are implemented. The program code can be executed entirely on the machine, partially on the machine, as a stand-alone software package, partially on the machine and partially on a remote machine, or entirely on a remote machine or server.

[0133] In the context of the present invention, machine-readable medium can be a tangible medium that can contain or store a program for use with an instruction execution system, device or equipment or used in combination with an instruction execution system, device or equipment. Machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. Machine-readable medium can include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared or semiconductor systems, devices or equipment, or any suitable combination of the foregoing. More specific examples of machine-readable storage media can include electrical connections based on one or more lines, portable computer disks, hard disks, random access memories (RAM), read-only memories (ROM), erasable programmable read-only memories (EPROM or flash memory), optical fibers, portable compact disk read-only memories (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0134] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and such modifications or substitutions are intended to be within the scope of protection of the present invention. Therefore, the scope of protection of the present invention shall be subject to the scope of protection of the claims.

Claims

1. A clustering federated learning method based on model decomposition, characterized in that: The method is based on model decomposition and low-rank similarity implementation and includes the following steps: Step 1: Model decomposition and initialization: Split the client model into a shallow embedding model and a deep decision model, and initialize the server's global embedding parameters and cluster center parameters; Step 2: Based on the low-rank model similarity clustering algorithm, the client model splicing matrix is ​​decomposed into a low-rank matrix by dimensionality reduction technology, and the low-rank matrix similarity is calculated; Step 3: Knowledge integration and parameter update mechanism: Combine the global embedding knowledge and cluster center knowledge from step 1 to optimize and update the client model parameters; Step 4: Dynamic update of cluster assignment and cluster center model parameters: Dynamically adjust the cluster assignment matrix based on the low-rank representation of the low-rank matrix, and update the global embedding and cluster center parameters.

2. A clustering federated learning method based on model decomposition according to claim 1, characterized in that: In step 1, the shallow embedding model is responsible for processing the low-dimensional features of the input data and focusing on the global structure of the data; the deep decision model focuses on capturing device-specific data distribution and is suitable for personalized learning.

3. The clustering federated learning method based on model decomposition according to claim 1, characterized in that: The calculation process of the low-rank model similarity includes: The server samples the client's local model and selects some clients for offline calculation; Flatten and concatenate the local models of the selected clients into the matrix W d ; For the matrix W d Apply dimensionality reduction technology to perform low-rank decomposition and obtain a low-rank matrix M; Calculate the similarity between the client model and the cluster center, that is, the low-rank matrix similarity.

4. The clustering federated learning method based on model decomposition according to claim 1, characterized in that: The process of knowledge integration and parameter updating in step 3 includes: Server Aggregation Client Model w i Forming a global embedding g e ; The server calculates the kth cluster center model where r i (k) is the cluster assignment of the i-th client model at the k-th cluster center, and m is the number of client models; The server will globally embed g e and the corresponding cluster center model Send to each client; The client updates the local model parameters based on the received global embedding and cluster center model.

5. The clustering federated learning method based on model decomposition according to claim 4, characterized in that: The local model parameters are updated based on a triple optimization objective, wherein the triple optimization objective includes a supervision loss L1, an intra-cluster regularization loss L2, and a global regularization loss L3; Minimizing the sum of the three losses is used as the comprehensive optimization objective function for updating local model parameters.

6. A clustering federated learning method based on model decomposition according to claim 5, characterized in that: The supervised loss is based on the task-related loss of the client's local data to ensure the model's fit to the local data; The intra-cluster regularization loss is used to control the closeness between the client model and the model of the cluster center to which it belongs, and to maintain the consistency number with the cluster center; The global regularization loss is used to promote global knowledge sharing and ensure the consistency of shallow embedding models with the global embedding.

7. The clustering federated learning method based on model decomposition according to claim 1, characterized in that: The dynamic update of cluster center model parameters includes: For cluster center k, calculate the weighted average of all client models assigned to the cluster to update the cluster center model Among them, r i (k) For the i-th client model w i Cluster assignment at the kth cluster center, where m is the number of client models; Based on shallow embedding model Update the global embedding g e : The updated cluster center model and global embedding are broadcast to the corresponding clients for the next round of local training.

8. The clustering federated learning method based on model decomposition according to claim 7, characterized in that: The method further includes updating the local model on the client side, the process of which includes: The client receives the global embedding g sent by the server e and cluster center model Perform multiple rounds of local training, updating the client model in each round according to the following formula: Among them, w i 'For the updated i-th client model, is the supervised loss gradient, μ is the intra-cluster regularization parameter, λ is the global regularization parameter, and η is the learning rate; The updated client model w i and low-rank representation M·w i Send back to the server, where M is a low-rank matrix.

9. The clustering federated learning method based on model decomposition according to claim 1, characterized in that: The updating of the cluster allocation includes: calculating the similarity matrix between the client and each cluster center based on the low-rank model similarity; and updating the cluster allocation matrix according to the principle of maximum similarity.

10. An electronic device comprising a memory and a processor, wherein a computer program is stored in the memory, wherein: When the processor executes the program, the method according to any one of claims 1 to 9 is implemented.

Images