Federal learning method and system for model heterogeneous scene

Abstract

The invention provides a federal learning method and system for a model heterogeneous scene, and the method comprises the steps: S1, collecting an individualized prototype, and enhancing an adaptive boundary; s2, optimizing the individualized prototype through comparative learning; s3, training of a local model is guided through the individualized prototype, and a client model is obtained; the individualized prototype is a prototype vector with client information. According to the method, only a low-dimensional class prototype is transmitted, so that the communication overhead is remarkably reduced, and specifically, the communication overhead provided by the invention is only 4% of that of a method of sharing a lightweight auxiliary model.

Description

Technical Field

[0001] The present invention belongs to the field of federated learning technology, and in particular, relates to a federated learning method and system for model heterogeneity scenarios. More specifically, it is a federated learning method for processing data and model heterogeneity. Background Art

[0002] With the rapid development of big data and artificial intelligence technology, data-driven machine learning models have been widely used in various fields. However, the issue of data privacy protection has become increasingly important, especially in fields involving sensitive information such as medicine and finance.

[0003] Federated Learning, or FL for short, is a new distributed learning paradigm that enables multi-party collaborative machine learning models while protecting data privacy by allowing each data owner to train the model locally and only share model parameters.

[0004] However, one of the main challenges faced by federated learning in practical applications is the heterogeneity of data and models. There is significant heterogeneity in the data sets held by different data owners, which is manifested in many aspects such as feature dimensions, data distribution, and sample size. Traditional information aggregation methods often rely on a unified data set and are difficult to effectively address the challenges of heterogeneous data in a federated learning environment. This information heterogeneity is more obvious when the client model structures participating in federated learning are heterogeneous. At the same time, model heterogeneity also makes the previous common model parameter aggregation method invalid, and a new information aggregation method independent of model parameters is needed.

[0005] Existing methods generally adopt the method of feature prototype sharing and averaging, and use the obtained global prototype as shared information to empower clients. However, simply averaging the feature prototypes uploaded by the client can easily lead to negative impacts between feature prototypes with heterogeneous information, and eliminating this negative impact is crucial.

[0006] In addition, in federated learning, each round of iteration involves the transmission of model parameters between a global server and multiple local clients. Communication-efficient federated learning can significantly reduce network communication costs by reducing the frequency of model updates, compressing the amount of transmitted data, or adopting more efficient encoding and transmission techniques. This is particularly beneficial for clients that need to connect via low-bandwidth networks, such as devices in remote areas or IoT devices.

[0007] Patent document CN119006895A discloses a personalized federated learning method for image classification, which includes: step 1: initialize the global prototype based on the server step 2: send the initialized global prototype to the client; step 3: the client initializes the local model parameters to randomly divide the local data set, update the global prototype, and obtain the updated model parameters and local prototype; step 4: the client retains the updated model parameters and uploads the updated local prototype to the server; step 5: the server aggregates the local prototypes uploaded by all clients to generate a new global prototype; step 6: repeat steps 2-6, send the global prototype obtained each time to the client for updating, until the maximum number of cycles is reached, and output the final global prototype. The present invention can better alleviate the impact of data heterogeneity on model accuracy, obtain higher classification accuracy, and greatly reduce the communication cost between the client and the server.

[0008] The defects of patent document CN119006895A are: this method trains local prototypes locally on the client, which increases the computational and time overhead of the client when participating in federated learning, but only uses weighted average aggregation of local prototypes to perform global prototype generation operations on the server side, wasting the server's computing resources; this method only updates the prototype locally, so that the local prototype can only learn local data information; this method requires that the local data set be divided into a meta-training set and a meta-test set, and has high requirements on the number of samples in the local data set; although this method is a personalized federated learning method, the global prototype is used for the boot process of local model training, and the local prototype obtained through local training is also for the purpose of aggregating a unified global prototype on the server side.

[0009] This problem needs to be solved urgently. Summary of the invention

[0010] In view of the defects in the prior art, the object of the present invention is to provide a federated learning method and system for model heterogeneous scenarios.

[0011] A federated learning method for model heterogeneity scenarios provided by the present invention includes:

[0012] Step S1: Collect individualized prototypes and enhance adaptive boundaries;

[0013] Step S2: Optimizing the individualized prototype by contrastive learning;

[0014] Step S3: guiding the training of the local model through the individualized prototype to obtain the client model;

[0015] The individualized prototype is a prototype vector having client information.

[0016] Preferably, in step S1, the personalized prototype is trained by a loss function to enhance the adaptive boundary; the mathematical expression of the loss function is:

[0017]

[0018] in, represents the loss function used to update the cth individualized prototype of the ith client; I t is the set of clients participating in the tth round; φ represents the Euclidean distance, and δ(t) represents the adaptive boundary; Represents a personalized prototype; represents the personalized prototype of the i-th client; the symbol ′ represents another one; The prototype representing the cth local prototype of the ith client;

[0019] The mathematical expression of the adaptive boundary is:

[0020]

[0021] in, represents the client prototype center of each category, τ is a threshold; C represents the total number of prototypes uploaded by each client; and Consistent, where c′ represents another category different from c; c represents a category, i.e., a prototype;

[0022] Said The mathematical expression is:

[0023]

[0024] Among them, I t represents the randomly sampled client subset in the tth iteration, that is, the set of clients participating in the tth round;

[0025] Said The mathematical expression is:

[0026]

[0027] in, represents the data subset belonging to category c in client i; Expressing hope, represents the data subset belonging to category c in client i; f i represents the feature extractor of client i; x represents any local data; θ i Represents the local feature extractor model parameters.

[0028] Preferably, in step S2, the loss function corresponding to each individualized prototype is minimized, the gradient is obtained, and the model is updated, wherein the mathematical expression of the learning rate is:

[0029]

[0030] Among them, η s represents the server learning rate; represents the derivative symbol;

[0031] Preferably, in step S3, the mathematical expression of the loss function of the client model is:

[0032]

[0033] Among them, e is the cross entropy loss function, w i are the local classifier model parameters, h i is a local classifier; λ represents a hyperparameter, E represents expectation; φ represents the Euclidean distance; represents the personalized prototype of the i-th client; y represents the label.

[0034] A federated learning system for model heterogeneity scenarios provided by the present invention includes:

[0035] Module M1: Collect individualized prototypes and enhance adaptive boundaries;

[0036] Module M2: Optimizing the individualized prototype through contrastive learning;

[0037] Module M3: guiding the training of the local model through the individualized prototype to obtain the client model;

[0038] The individualized prototype is a prototype vector having client information.

[0039] Preferably, in the module M1, the personalized prototype is trained by a loss function to enhance the adaptive boundary; the mathematical expression of the loss function is:

[0040]

[0041] in, represents the loss function used to update the cth individualized prototype of the ith client; I t is the set of clients participating in the tth round; φ represents the Euclidean distance, and δ(t) represents the adaptive boundary; Represents a personalized prototype; represents the personalized prototype of the i-th client; the symbol ′ represents another one; The prototype representing the cth local prototype of the ith client;

[0042] The mathematical expression of the adaptive boundary is:

[0043]

[0044] in, represents the client prototype center of each category, τ is a threshold; C represents the total number of prototypes uploaded by each client; and Consistent, where c′ represents another category different from c; c represents a category, i.e., a prototype;

[0045] Said The mathematical expression is:

[0046]

[0047] Among them, I t represents the randomly sampled client subset in the tth iteration, that is, the set of clients participating in the tth round;

[0048] Said The mathematical expression is:

[0049]

[0050] in, represents the data subset belonging to category c in client i; Expressing hope, represents the data subset belonging to category c in client i; f i represents the feature extractor of client i; x represents any local data; θ i Represents the local feature extractor model parameters.

[0051] Preferably, in the module M2, the loss function corresponding to each individualized prototype is minimized, the gradient is obtained, and the model is updated, wherein the mathematical expression of the learning rate is:

[0052]

[0053] Among them, η s represents the server learning rate; Represents the derivative symbol.

[0054] Preferably, in the module M3, the mathematical expression of the loss function of the client model is:

[0055]

[0056] Among them, l is the cross entropy loss function, w i are the local classifier model parameters, h iis a local classifier; λ represents a hyperparameter, E represents expectation; φ represents the Euclidean distance; represents the personalized prototype of the i-th client; y represents the label.

[0057] According to a computer-readable storage medium storing a computer program provided by the present invention, when the computer program is executed by a processor, the steps of the federated learning method for model heterogeneous scenarios are implemented.

[0058] An electronic device provided according to the present invention includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is executed by the processor, the steps of the federated learning method for model heterogeneous scenarios are implemented.

[0059] Compared with the prior art, the present invention has the following beneficial effects:

[0060] 1. The present invention significantly reduces communication overhead by transmitting only low-dimensional class prototypes. Specifically, the communication overhead provided by the present invention is only 4% of that of methods such as sharing lightweight auxiliary models.

[0061] 2. The present invention improves the performance of the model through trainable personalized prototypes and adaptive boundary enhanced contrast learning, and ensures extremely low communication efficiency.

[0062] 3. FedpTGP provided by the present invention outperforms all baseline methods on four data sets, with the highest accuracy improvement reaching 9.89%.

[0063] 4. The present invention directly uses local data to train the model; the present invention converts the local prototype into a personalized trainable prototype on the server side and refers to the information of other local prototypes, so that the personalized trainable prototype obtained according to the present invention can not only maintain the original heterogeneity to meet the local feature information of each client, but also absorb the feature information from other clients. In addition, through adaptive boundary enhanced contrast learning, the present invention can amplify this effect and improve the inter-class distinguishability of the personalized trainable prototype, thereby improving its quality.

[0064] 5. The present invention does not perform additional training locally, but instead performs the optimization and update of the local prototype on the server, which can reduce client overhead and make full use of the server's computing resources. The present invention trains the local prototype on the server, and in the process of updating the local prototype, refers to the features of local prototypes from other clients, and uses adaptive boundary enhancement contrast learning and collaborative optimization to enable the trained personalized trainable prototype to absorb information from other clients, thereby producing a better training guidance effect on the local client. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] Other features, objects and advantages of the present invention will become more apparent from the detailed description of non-limiting embodiments made with reference to the following drawings:

[0066] Figure 1 Schematic diagram of the heterogeneous federated learning architecture provided by the present invention. DETAILED DESCRIPTION

[0067] The present invention is described in detail below in conjunction with specific embodiments. The following embodiments will help those skilled in the art to further understand the present invention, but are not intended to limit the present invention in any form. It should be noted that, for those of ordinary skill in the art, several changes and improvements can also be made without departing from the concept of the present invention. These all belong to the protection scope of the present invention.

[0068] like Figure 1 The following is a diagram of the heterogeneous federated learning architecture. Specifically, different participating clients use models with different architectures.

[0069] The present invention relates to a federated learning method based on trainable personalized prototypes and adaptive boundary-enhanced contrastive learning. The method solves the challenges brought by data and model heterogeneity in federated learning by introducing trainable personalized prototypes and adaptive boundary-enhanced contrastive learning.

[0070] Specifically, each client calculates the prototype of each category from its dataset and sends it to the server, which uses these prototypes to train personalized prototypes and optimizes them through adaptive boundary enhanced contrastive learning. The optimized personalized prototypes are then sent back to the client to guide the training of the client model and improve the separability of feature representation.

[0071] According to the federated learning method for model heterogeneous scenarios provided by the present invention, the process of the software system thereof includes:

[0072] Step S1: Collect individualized prototypes and enhance adaptive boundaries;

[0073] Step S2: Optimizing the individualized prototype by contrastive learning;

[0074] Step S3: guiding the training of the local model through the individualized prototype to obtain the client model;

[0075] Specifically, the individualized prototype is a prototype vector having client-specific information, which is information that distinguishes the client from other clients.

[0076] The prototype vector, whose English name is prototype, is a representative feature vector in each type of data.

[0077] In the step S1, the adaptive boundary of the personalized prototype is enhanced as follows:

[0078] In step S1, the server is instructed to train the personalized prototype using the following loss function, the mathematical expression of which is:

[0079]

[0080] Among them, I t is the set of clients participating in round t, φ represents the Euclidean distance, and δ(t) is the adaptive boundary; represents the loss function used to update the cth individualized prototype of the ith client; Represents a personalized prototype;

[0081] The mathematical expression of the adaptive boundary is:

[0082]

[0083] Among them, δ(t) represents the adaptive boundary; represents the prototype center of each category of the client, τ is a threshold to prevent the boundary from growing infinitely; φ represents the Euclidean distance; c represents any prototype of the client; the symbol ′ represents another; C represents the total number of prototypes uploaded by each client; and Consistent, where c′ represents another category different from c; c represents a category, i.e., a prototype;

[0084] Said The mathematical expression is:

[0085]

[0086] Among them, I t represents the randomly sampled client subset in the tth iteration, that is, the set of clients participating in the tth round;

[0087] Said The mathematical expression is:

[0088]

[0089] in, represents the data subset belonging to category c in client i; represents the prototype of the cth local prototype of the ith client; E represents the expectation, represents the data subset belonging to category c in client i; f i Represents the client

[0090] feature extractor of i; x represents any local data; θ irepresents the local feature extractor model parameters;

[0091] In the step S2, the contrastive learning optimizes the individualized prototype, which is specifically achieved as follows:

[0092] Minimize the corresponding The loss function obtains the gradient and updates the model: is the server learning rate;

[0093] In the step S3, the training of the local model is guided by the individualized prototype to obtain the target model of the client;

[0094] This is achieved by minimizing the local client loss L i :

[0095]

[0096] In step S3, the client model is trained, and the mathematical expression of the loss function of the client model is:

[0097]

[0098] Among them, l is the cross entropy loss function, w i are the local classifier model parameters, h i is a local classifier; λ represents a hyperparameter, E represents expectation; φ represents the Euclidean distance; Represents the personalized prototype of the i-th client.

[0099] The method is implemented by transferring only low-dimensional class prototypes.

[0100] In the model heterogeneous scenario considered by the present invention, the feature extraction capabilities of the models vary greatly, and the local prototypes obtained through local training are also heterogeneous between different clients, causing the global model finally aggregated to lose its unified representation capability.

[0101] A federated learning method based on trainable personalized prototypes and adaptive boundary enhanced contrastive learning provided by the present invention includes: each client calculates a prototype of each category from its data set, namely, a personalized prototype; the client sends the calculated prototype to a server; after receiving the prototypes of all clients, the server uses trainable personalized prototypes and adaptive boundary enhanced contrastive learning to optimize the personalized prototype; the server sends the trained personalized prototype back to the client; the client uses the personalized prototype to guide the training of the local model to improve the separability of feature representation by minimizing a specific loss function.

[0102] The personalized prototype is further processed through a neural network model to improve its training ability.

[0103] The adaptive boundary δ(t) is defined as:

[0104]

[0105] in, represents the client prototype center of each category, and τ is a threshold. in represents the data subset belonging to category c in client i.

[0106] The loss function of the client model is defined as:

[0107]

[0108] The method is implemented by transferring only low-dimensional class prototypes.

[0109] 1. Initialization: Assume there are M clients, each with a different data set and their respective heterogeneous model architectures. Each client’s model is divided into a feature extractor f i and classifier h i Initialize a trainable personalized prototype for each client i, i.e. and a unified further processing model F.

[0110] 2. Iteration process:

[0111] Server randomly samples clients: The server randomly samples a subset of clients in the tth iteration, that is, I t .

[0112] Server sends personalized prototype: The server sends the personalized prototype Sent to a subset of clients, namely I t .

[0113] Client updates the model: The client uses the personalized prototype to guide the training of the local model and updates the model parameters.

[0114] Client computes prototypes: The client computes the prototype of each category from its dataset, i.e.

[0115] Client sends prototype: The client sends the calculated prototype to the server.

[0116] Server calculates adaptive bounds: The server calculates the adaptive bound, δ(t), based on the client prototype.

[0117] Server updates personalized prototype: The server updates the personalized prototype using adaptive boundary-enhanced contrastive learning

[0118] 3. Evaluation method:

[0119] Datasets: We conduct experiments using four popular datasets, including CIFAR-10, CIFAR-100, Flowers-102, and Tiny-ImageNet.

[0120] Baseline methods: The present invention is compared with six existing federated learning methods, including LG-FedAvg, FedGen, FML, FedKD, FedDistill and FedProto.

[0121] Model Heterogeneity: Unless otherwise specified, heterogeneous feature extractors, i.e., HtFE settings, are evaluated. We use “HtFEX” to denote HtFE settings, where X is the number of different model architectures in federated learning. We assign the i-th client to the (imodX)-th model architecture.

[0122] Statistical Heterogeneity: We conduct experiments under two widely used statistical heterogeneity settings: a pathological setting and a realistic setting.

[0123] Evaluation results,performance comparison: The experimental results show that the proposed FedpTGP outperforms all baseline methods on four data sets, with the highest accuracy improvement reaching 9.89%.

[0124] Impact of model heterogeneity: The present invention performs well under different model heterogeneity settings, with the highest improvement reaching 6.31%.

[0125] like Figure 1 As shown in the figure, D i represents the dataset of client i, f i represents the feature extractor of client i, h i represents the classifier of client i; D j represents the dataset of client j, f j represents the feature extractor of client j, h j represents the classifier of client j.

[0126] In other words, in order to give full play to the information contained in the client prototype, the present invention proposes a federated learning method based on trainable personalized prototypes and adaptive boundary enhanced contrastive learning, namely FedpTGP, which includes:

[0127] System architecture: Client: Each client has its own dataset and heterogeneous model architecture. The client model is divided into feature extractor f iand classifier h i . Server: The server is responsible for receiving the prototype from the client, training the personalized prototype, and sending the personalized prototype back to the client.

[0128] Client model training:

[0129] Each client i first computes the prototype of each category from its dataset in represents the data subset belonging to category c in client i.

[0130] The client sends the calculated prototype to the server.

[0131] Personalized Prototype Training:

[0132] After the server receives the prototypes of all clients, it uses the trainable personalized prototype And adaptive boundary enhanced contrastive learning, namely ACL, to optimize the personalized prototype. Among them, the personalized prototype It is further processed by a neural network model F to improve its training ability.

[0133] The server trains the personalized prototype using the following loss function:

[0134]

[0135] Among them, I t is the set of clients participating in round t, φ represents the Euclidean distance, and δ(t) is the adaptive boundary, which is defined as:

[0136]

[0137] in, Represents the client prototype center for each category, yes In the set of round t, τ is a threshold to prevent the boundary from growing infinitely.

[0138] Client model updates:

[0139] The server will train the personalized prototype Send back to the client.

[0140] The client uses the personalized prototype to guide the training of the local model to improve the separation of feature representations by minimizing the following loss function:

[0141]

[0142] The present invention also provides a federated learning system for model heterogeneous scenarios. The federated learning system for model heterogeneous scenarios can be implemented by executing the process steps of the federated learning method for model heterogeneous scenarios, that is, those skilled in the art can understand the federated learning method for model heterogeneous scenarios as a preferred implementation of the federated learning system for model heterogeneous scenarios.

[0143] Those skilled in the art know that, in addition to realizing the system and its various devices, modules, and units provided by the present invention in a purely computer-readable program code, it is entirely possible to realize the same functions in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers by logically programming the method steps. Therefore, the system and its various devices, modules, and units provided by the present invention can be considered as a hardware component, and the devices, modules, and units included therein for realizing various functions can also be regarded as structures within the hardware component; the devices, modules, and units for realizing various functions can also be regarded as both software modules for realizing the method and structures within the hardware component.

[0144] In the description of the present application, it should be understood that the terms "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside", etc., indicating orientations or positional relationships, are based on the orientations or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the referred device or element must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be understood as a limitation on the present application.

[0145] The above describes the specific embodiments of the present invention. It should be understood that the present invention is not limited to the above specific embodiments, and those skilled in the art can make various changes or modifications within the scope of the claims, which does not affect the essence of the present invention. In the absence of conflict, the embodiments of the present application and the features in the embodiments can be combined with each other arbitrarily.

Claims

1. A federated learning method for model heterogeneity scenarios, characterized in that: include: Step S1: Collect individualized prototypes and enhance adaptive boundaries; Step S2: Optimizing the individualized prototype by contrastive learning; Step S3: guiding the training of the local model through the individualized prototype to obtain the client model; The individualized prototype is a prototype vector having client information.

2. The federated learning method for model heterogeneous scenarios according to claim 1, characterized in that: In step S1, the personalized prototype is trained by a loss function to enhance the adaptive boundary; the mathematical expression of the loss function is: in, represents the loss function used to update the cth individualized prototype of the ith client; I t is the set of clients participating in the tth round; φ represents the Euclidean distance, and δ(t) represents the adaptive boundary; Represents a personalized prototype; represents the personalized prototype of the i-th client; the symbol ′ represents another one; The prototype representing the cth local prototype of the ith client; The mathematical expression of the adaptive boundary is: in, represents the client prototype center of each category, τ is a threshold; C represents the total number of prototypes uploaded by each client; and Consistent, where c′ represents another category different from c; c represents a category, i.e., a prototype; Said The mathematical expression is: Among them, I t represents the randomly sampled client subset in the tth iteration, that is, the set of clients participating in the tth round; Said The mathematical expression is: in, represents the data subset belonging to category c in client i; E represents the expectation, represents the data subset belonging to category c in client i; f i represents the feature extractor of client i; x represents any local data; θ i Represents the local feature extractor model parameters.

3. The federated learning method for model heterogeneous scenarios according to claim 2, characterized in that: In step S2, the loss function corresponding to each individualized prototype is minimized, the gradient is obtained, and the model is updated, wherein the mathematical expression of the learning rate is: Among them, η s represents the server learning rate; Represents the derivative symbol.

4. The federated learning method for model heterogeneous scenarios according to claim 3, characterized in that: In step S3, the mathematical expression of the loss function of the client model is: Among them, l is the cross entropy loss function, w i are the local classifier model parameters, h i is a local classifier; λ represents a hyperparameter, E represents expectation; φ represents the Euclidean distance; represents the personalized prototype of the i-th client; y represents the label.

5. A federated learning system for model heterogeneity scenarios, characterized in that: include: Module M1: Collect individualized prototypes and enhance adaptive boundaries; Module M2: Optimizing the individualized prototype through contrastive learning; Module M3: guiding the training of the local model through the individualized prototype to obtain the client model; The individualized prototype is a prototype vector having client information.

6. The federated learning method for model heterogeneous scenarios according to claim 5, characterized in that: In the module M1, the personalized prototype is trained by the loss function to enhance the adaptive boundary; the mathematical expression of the loss function is: in, represents the loss function used to update the cth individualized prototype of the ith client; I t is the set of clients participating in the tth round; φ represents the Euclidean distance, and δ(t) represents the adaptive boundary; Represents a personalized prototype; represents the personalized prototype of the i-th client; the symbol ′ represents another one; The prototype representing the cth local prototype of the ith client; The mathematical expression of the adaptive boundary is: in, represents the client prototype center of each category, τ is a threshold; C represents the total number of prototypes uploaded by each client; and Consistent, where c′ represents another category different from c; c represents a category, i.e., a prototype; Said The mathematical expression is: Among them, I t represents the randomly sampled client subset in the tth iteration, that is, the set of clients participating in the tth round; Said The mathematical expression is: in, represents the data subset belonging to category c in client i; E represents the expectation, represents the data subset belonging to category c in client i; f i represents the feature extractor of client i; x represents any local data; θ i Represents the local feature extractor model parameters.

7. The federated learning system for model heterogeneous scenarios according to claim 6, characterized in that: In the module M2, the loss function corresponding to each individualized prototype is minimized, the gradient is obtained, and the model is updated, where the mathematical expression of the learning rate is: Among them, η s represents the server learning rate; Represents the derivative symbol.

8. The federated learning system for model heterogeneity scenarios according to claim 7, characterized in that: In the module M3, the mathematical expression of the loss function of the client model is: Among them, l is the cross entropy loss function, w i are the local classifier model parameters, h i is a local classifier; λ represents a hyperparameter, E represents expectation; φ represents the Euclidean distance; represents the personalized prototype of the i-th client; y represents the label.

9. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the steps of the federated learning method for model heterogeneous scenarios described in any one of claims 1 to 4 are implemented.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the computer program is executed by a processor, the steps of the federated learning method for model heterogeneous scenarios described in any one of claims 1 to 4 are implemented.

Images