Federal learning method capable of filtering noise label samples and related device

Abstract

The invention belongs to the technical field of artificial intelligence, and discloses a federal learning method capable of filtering noise label samples and a related device. The federal learning method comprises the steps that a selected client downloads parameters of a global model from a server; training a local personalized model and a local global model by using locally selected non-noise label samples, and updating and uploading parameters of the locally trained global model to a server during training; the personalized model is used for cleaning local data to select non-noise label sample data for local training; the local personalized model and the global model select samples from each other and are alternately updated; and the server aggregates the parameter updates uploaded by the clients. According to the technical scheme of the invention, for a scene in which local data label noise distribution of each client is inconsistent, dynamic filtering of noise labels and improvement of model stability and generalization performance are realized.

Description

Technical Field

[0001] The present invention belongs to the field of artificial intelligence technology, and in particular relates to a federated learning method and related devices capable of filtering noise label samples. Background Art

[0002] With the development of computer technology, major breakthroughs have occurred in the fields of artificial intelligence and machine learning. Along with technological development, a large amount of data has been generated, including sensitive privacy data such as medical and financial data. Therefore, the data security of personal privacy data has become an increasing focus of attention.

[0003] Against this backdrop, federated learning, a distributed learning method, emerged. It allows for learning common models from different clients while ensuring the privacy of each client's data, thus addressing the issue of data privacy protection to a certain extent. The performance of federated learning depends largely on the quality of data sample labels. In practical applications, the data distribution between clients is often different, and the quality of the data labels held by each client also varies. Label errors (i.e., label noise) in these differently distributed data can significantly impact the generalization ability of the federated learning model. It is important to emphasize that, unlike the noisy label problem in centralized learning, federated learning faces two special circumstances: 1) Human-induced label errors in each client can lead to different noise distributions for each client's labels, i.e., different label errors between clients; and 2) the limited number of samples held by each client can easily lead to overfitting during local training. Given the above situation, many current methods that can solve noisy labels in centralized learning cannot be directly applied to federated learning. These methods for solving label noise in centralized learning will encounter huge challenges in the federated learning context. Moreover, the server aggregation method in federated learning has difficulty learning samples unique to the client, which may lead to the inability to determine whether the samples that cause high loss function values are noise or samples unique to the client.

[0004] Currently, existing label noise processing methods for federated learning generally ignore the problem of different distributions of noisy label samples in federated learning. These methods encounter difficulties in identifying noisy label samples in different clients. Among them, some methods require the assumption that some clients hold samples without label noise, and cannot be applied to scenarios where all clients have label noise, making it difficult to achieve the desired goal. Summary of the Invention

[0005] The present invention aims to provide a federated learning method and related apparatus capable of filtering noisy label samples, thereby addressing one or more of the aforementioned technical problems. The disclosed technical solution, designed for scenarios where the local data label noise distribution of each client is inconsistent, achieves dynamic filtering of noisy labels and improves model stability and generalization performance.

[0006] In order to achieve the above object, the present invention adopts the following technical solutions:

[0007] In a first aspect, the present invention provides a federated learning method capable of filtering noisy label samples, comprising the following steps:

[0008] In each round of federated learning training, the selected client downloads the parameters of the global model from the server;

[0009] For each selected client, the local personalized model and global model are trained using locally selected non-noise labeled samples. Upon completion of training, the parameters of the locally trained global model are updated and uploaded to the server. The personalized model is used to clean local data to select non-noise labeled sample data for local training. During training, the local personalized model and the global model are updated alternately by selecting samples.

[0010] The server aggregates the parameter updates uploaded by each client and starts the next round of federated learning.

[0011] A further improvement of the technical solution of the present invention is that in the step of using locally selected non-noise label samples to train the local personalized model and the global model, the local personalized model of the client and the global model are trained collaboratively, the number of training iterations of the two models is the same, and training and parameter updates are performed in each iteration.

[0012] A further improvement of the technical solution of the present invention is that the step of using the locally selected non-noise label samples to train the local personalized model and the global model specifically includes:

[0013] The sample selector is used to calculate and select non-noise label samples, and the loss function values of the client's local global model and personalized model on the non-noise label samples are calculated respectively. The parameters of the client's local global model and personalized model are calculated and updated.

[0014] A further improvement of the technical solution of the present invention is that in the step of calculating and selecting non-noise label samples by the sample selector,

[0015] The formula for selecting non-noise label samples is expressed as:

[0016]

[0017] Where, v n Indicates whether the example n is a noise label sample, v n =1, then example n is a non-noise label sample, v n = 0, then example n is a noise label sample; l CE (·) is the cross entropy loss function; L is the category of the label; x n is the characteristic of example n; is the label of example n; f(·) is the logistic regression classifier;

[0018] The parameters of the logistic regression classifier f(·) are calculated using the following formula:

[0019]

[0020] Where, is a given function space; n k is the data sample selected by client k, is the mean number of data samples selected by client k;

[0021]

[0022] Where β is a predetermined hyperparameter; For the conditional distribution Lower your expectations, For a given noise dataset and noise labels The probability distribution of .

[0023] A further improvement of the technical solution of the present invention is that in the step of respectively calculating the loss function values of the client's local global model and the personalized model on the non-noise label samples,

[0024] First, initialize the client's local personalized model parameters. The initialized parameters are consistent with the parameters of the global model downloaded by the client from the server.

[0025] The objective formula F(·) for global model training is as follows:

[0026]

[0027] Where w is the parameter of the global model; is the total number of samples selected on all clients; F k (·) represents the target formula during local client training, which is used to characterize the prediction loss of local data; K is the total number of clients;

[0028] F k (·)The calculation formula is as follows:

[0029]

[0030] Where, v n Indicates whether the example n is a noise label sample, l(·) is the loss function, x n is the feature of example n, is the label of example n, and λ represents the parameter that controls the influence of the global model on the personalized model.

[0031] A further improvement of the technical solution of the present invention is that in the step of calculating and updating the parameters of the local global model and the personalized model of the client,

[0032] The parameter update method of the personalized model in the client is expressed as:

[0033]

[0034] The parameter update method of the global model in the client is expressed as:

[0035] w k ←w k -ηλ(w k -θ(w k ));

[0036]

[0037] Where θ k is the personalized model of client k; w k is the parameter weight of the model held by client k; ζ is the local learning rate; is the gradient operator; l k (·) is the loss function of client k; η is the global learning rate; θ is the global model.

[0038] A further improvement of the technical solution of the present invention is that, in the step of the server aggregating parameter updates uploaded by each client, the calculation expression of the aggregation parameter operation is:

[0039]

[0040] Where, represents the model parameter update uploaded by client k in round t+1, where t represents the round number; γ∈[0,1] is a hyperparameter. The larger γ is, the more the server relies on the parameter update uploaded by the client when aggregating the client uploaded parameters and updating the global model parameters; w t is the global model parameter of server t round, w t+1 are the global model parameters of the server in round t+1.

[0041] In a second aspect of the present invention, an electronic device is provided, comprising a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the program, the federated learning method for filtering noise label samples as described in any one of the first aspects of the present invention is implemented.

[0042] In a third aspect of the present invention, a non-transitory computer-readable storage medium is provided, on which a computer program is stored, and when the computer program is executed by a processor, the federated learning method for filtering noise label samples as described in any one of the first aspects of the present invention is implemented.

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

[0044] The present invention specifically discloses a federated learning method that can filter noisy label samples. By designing a dual-model structure on the client, it realizes the recognition and filtering of noisy label data with different distributions possessed by each client, and solves the challenges brought by different category distributions and human labeling errors. Specifically and explanatorily, the dual models loaded on the client include a global model and a personalized model. During the local training process of the client, the two models select samples from each other and update them alternately. Compared with the training of a single model, this method of the present invention reduces the risk of accumulated errors caused by noise data during the training process.

[0045] Furthermore, in order to reduce the overfitting of the local model, the technical solution of the present invention also designs and implements two regularizers to constrain the model. Specifically, CR (·) is a regularizer designed in the method of the present invention to solve the overfitting problem of the local personalized model; The second regularizer designed for the present method is used to mitigate overfitting during local training. This regularizer reduces overfitting of the personalized model by constraining the distance between the local personalized model and the global model. Compared to existing noisy label processing methods for centralized learning, the present invention can address the model overfitting problem caused by local noisy label data.

[0046] In summary, the technical solution of the present invention can alleviate the impact of different noise distributions of sample labels between clients on the performance of local and global models through the technical means of dual-model collaborative training and regularization constraints. Explanatoryly, the technical solution of the present invention designs a dual-model structure at the node serving as the client, combining the local personalized model with the global model to effectively select non-noise label samples owned by the client. In addition, the technical solution of the present invention designs two regularizers to reduce the overfitting problem caused by local client training. The technical solution of the present invention has performed well in tests on various data sets and is suitable for fields such as medicine and finance that have high requirements for data privacy and label quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below; obviously, the drawings described below are some embodiments of the present invention, and for ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0048] Figure 1 1 is a flow chart of a federated learning method capable of filtering noisy label samples in an embodiment of the present invention;

[0049] Figure 2 is a schematic diagram of a framework of a federated learning method capable of filtering noisy label samples in an embodiment of the present invention;

[0050] Figure 3 Schematic diagram of the federated learning process between the server and the client in an embodiment of the present invention. DETAILED DESCRIPTION

[0051] In order to make the purpose, technical solutions and advantages of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention; it is obvious that the described embodiments and technical solutions are only part of the embodiments of the present invention, not all of the embodiments.

[0052] All other embodiments obtained by persons of ordinary skill in the art based on the technical solutions disclosed in the embodiments of the present invention without creative effort shall fall within the scope of protection of the present invention. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units that are not explicitly listed or that are inherent to these processes, methods, products, or apparatuses.

[0053] See also Figure 1 , an embodiment of the present invention provides a federated learning method capable of filtering noisy label samples, comprising the following steps:

[0054] Step 1: In each round of federated learning training, the selected client downloads the parameters of the global model from the server;

[0055] Step 2: For each selected client, the local personalized model and global model are trained using locally selected non-noise labeled samples. Upon completion of the training, the parameters of the locally trained global model are updated and uploaded to the server. The personalized model is used to clean the local data to select non-noise labeled sample data for local training. During the training process, the local personalized model and the global model are updated alternately by selecting samples.

[0056] Step 3: The server aggregates the parameter updates uploaded by each client and starts the next round of federated learning.

[0057] The technical solution disclosed in the embodiments of this invention utilizes a dual-model architecture designed on the client side. This architecture can better adapt to noisy label data with varying distributions, thereby enabling the identification and filtering of noisy label data with varying distributions across clients. This, in turn, addresses the challenges presented by heterogeneous category distributions and human labeling errors. Compared to other existing methods for processing noisy label data for federated learning, the technical solution of this invention can be applied to situations where each client has noisy label data with varying distributions, enabling better data cleaning and filtering.

[0058] See also Figure 2 and Figure 3 In an embodiment of the present invention, a federated learning method capable of filtering noise label samples is provided, comprising the following steps:

[0059] S1. The method of the present invention is applicable to scenarios where a server and K clients coordinate to jointly perform model training; wherein the server loads the global model, and these parameters can be called by the server's processor for subsequent calculations; the client loads the local personalized model and the global model downloaded from the server.

[0060] Initially, the server initializes the global model parameters for subsequent training. During each round of federated learning, the server specifies the clients participating in that round. After the client completes local training, the server receives the parameter updates returned by the client and aggregates them to obtain the new global model parameters.

[0061] The total training data set is a data set D of L-class labels of size N. These data contain features X and labels Y. These data are held by each client and stored on the client's local storage medium. The data set held by the client is D k , the union of the data sets held by each client is the total data set D, that is

[0062] The label noise held by each client k can be expressed using the noise transformation matrix T k To express it, the matrix is defined as follows:

[0063]

[0064] Where P(·) represents the probability, is the noise label, that is, the wrong label, and Y is the true label.

[0065] When T k =T and D k =D, the label noise distributions between clients are the same; in other cases, the label noise distributions between clients are considered to be different.

[0066] After the server initializes the global model parameters and selects the clients to participate in federated learning, the selected clients download the global model from the server for subsequent training and other processes.

[0067] S2. Each selected client will use its own data for training. There are two models locally on the client: a personalized model and a global model downloaded from the server.

[0068] The client will divide the locally held dataset into multiple parts according to a certain batch size.

[0069] During each training round, the client initializes its personalized model using the global model parameters downloaded from the server. The parameters of these two models are identical at the beginning of this round of federated learning training, but as the client progresses through the training round, the parameters of these two models are updated differently. The personalized model is used to clean local data, selecting low-noise labeled sample data for local training. At this time, the client also trains and updates the global model parameters downloaded from the server. These updated parameters are then uploaded back to the server for aggregation and subsequent rounds of federated learning.

[0070] The number of iterations E for model training on each client is pre-specified. In each round of federated learning, the model on the client will be iteratively updated e times.

[0071] In a specific exemplary technical solution, in each round, the specific training process of the selected client is as follows:

[0072] S201. Calculate and select clean samples without noise through a sample selector. Sample selection is performed according to the following formula:

[0073]

[0074] Among them, v n Used to indicate whether the example n is a noise label sample, when v n =1, example n is a non-noise label sample; when vn = 0, the sample n is a noise label sample; l CE (·) is the cross entropy loss function; L is the category of the label; x n is the feature of example n, is the label of example n, and f(·) is the logistic regression classifier.

[0075] S202, respectively calculate the dual models on the client k, that is, the global model w k With personalized model θ k The loss function value on the selected data sample;

[0076] In a specific exemplary technical solution, the client loads two models, namely the personalized model θ k and the global model w k The loss function values on the data samples are calculated respectively, and the samples used for training are screened according to their respective loss function values to obtain low-noise label samples d w ,d θ .

[0077] Specifically, in step S202, the logistic regression classifier f(·) in the formula based on the clean sample data is selected, and its parameters are calculated by the following formula:

[0078]

[0079] in, is a given function space; n k is the data sample selected by client k, is the mean number of data samples selected by client k;

[0080] In the above calculation formula, l CR (·) is a regularizer designed in this method to solve the overfitting of the local personalized model. Its specific definition formula is as follows:

[0081]

[0082] Among them, β is a predetermined hyperparameter; For the conditional distribution Lower your expectations, For a given noise dataset and noise labels The probability distribution of l; subsequently l(·) will be used to refer to l CE (·)+l CR (·).

[0083] The predicted loss value of the client's local data is represented by the following calculation formula:

[0084]

[0085] Among them, F k (·) represents the target formula during local client training. The second regularizer designed for this method is used to mitigate overfitting during local training. This regularizer reduces overfitting of the personalized model by constraining the distance between the local personalized model and the global model. λ is a parameter that controls the influence of the global model on the personalized model. A smaller λ indicates that the personalized model focuses more on local data information. Conversely, a larger λ indicates that the personalized model focuses more on global model information.

[0086] S203: Calculate and update the parameters of the dual model. The personalized model parameters are updated in accordance with the following formula:

[0087]

[0088] Among them, ζ is the local learning rate, l k (·) is the loss function of client k.

[0089] The parameter update method of the global model in the client is based on the following formula:

[0090] w k ←w k -ηλ(w k -θ(w k ));

[0091]

[0092] Among them, η is the global learning rate.

[0093] Due to the aforementioned data screening and cleaning process, the number of low-noise labeled sample data used by the client for model training may differ from the number of sample data held locally by the client. Furthermore, since the parameters of the personalized and global models change with each iteration, the number of samples selected for training may vary from iteration to iteration. Therefore, after completing the parameter updates for the personalized model and the local global model, the client needs to calculate the mean number of data samples used in each iteration. This will be used to calculate the weight of each client when the server aggregates the parameter updates from each client.

[0094] S3. After step S2 is completed, the parameter updates of the locally trained global model are uploaded.

[0095] The client not only uploads the parameters of the locally trained global model, but also uploads the mean number of data samples used in each local iteration for subsequent calculations.

[0096] The objective formula of the server global model is as follows:

[0097]

[0098] Among them, w is the parameter of the global model, is the total number of samples selected on all clients, F k (·) represents the target formula during local client training.

[0099] S4. The server aggregates the parameter updates uploaded by the client and starts the next round of federated learning.

[0100] The server aggregates the parameter updates uploaded by each client according to the following formula:

[0101]

[0102] Where t represents the round; Indicates the model parameter update uploaded by client k in round t+1; is the mean number of data samples uploaded locally by client k for each iteration; γ∈[0,1] is a hyperparameter. The larger γ is, the more the server relies on the client-uploaded parameters when aggregating the client-uploaded parameters and updating the global model parameters.

[0103] After the global model parameters are aggregated and updated on the server side, the next round of federated learning will be carried out, and the above operations will be repeated for multiple rounds until the model converges.

[0104] Table 1 shows the accuracy (in %) of the present invention's classification tasks using independent and identically distributed (IID) and non-IID data under different levels of label noise using the MNIST and CIFAR-10 datasets. MNIST is a dataset of handwritten digits containing 70,000 black-and-white images of handwritten digits, each with a label, for a total of 10 label categories; CIFAR-10 is a dataset of color images containing 60,000 images, each with a label, for a total of 10 label categories. ρ represents the ratio of clients with noisy label samples, and τ represents the lower limit of the client noise level. Table 2 shows the comparative accuracy (in %) of the present invention and other federated learning frameworks for classification tasks under independent and identically distributed noisy label samples.

[0105] Tables 1 and 2 show the classification task accuracy of the present invention on the MNIST and CIFAR-10 datasets and the comparison with other methods; other methods include: FedAvg (MCMAHAN HB, MOORE E, RAMAGE D, et al. Communication-Efficient Learning of Deep Networks from Decentralized Data;proceedings of the International Conference on Artificial Intelligence and Statistics, F, 2016[C].), FedCorr (XU J, CHEN Z, QUEK TQS, et al. FedCorr: Multi-Stage Federated Learning for Label Noise Correction;proceedings of the 2022 IEEE / CVF Conference on Computer Vision and Pattern Recognition (CVPR), F18-24 June 2022, 2022[C].).

[0106] Table 1. Accuracy of the classification task using independent and identically distributed (IID) and non-IID data with different label noise levels using the MNIST and CIFAR-10 datasets.

[0107]

[0108] Table 2. Comparison of the accuracy of the proposed method and other federated learning frameworks in classification tasks under independent and identically distributed noisy label samples

[0109]

[0110] Based on Tables 1 and 2, it can be seen that the present invention performs well under different data sets, different label noise levels, and different client-held data distributions. At the same time, when the noise label level is high, the performance of the present invention is better than other federated learning methods.

[0111] In one embodiment of the present invention, a computer device is provided, comprising a processor and a memory, wherein the memory is used to store a computer program, the computer program includes program instructions, and the processor is used to execute the program instructions stored in the computer storage medium. The processor may be a central processing unit (CPU), or may be other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components, etc., which are the computing core and control core of the terminal, and are suitable for implementing one or more instructions, specifically suitable for loading and executing one or more instructions in a computer storage medium to implement a corresponding method flow or corresponding function; the processor described in the embodiment of the present invention can be used to perform operations of a federated learning method that can filter noise label samples.

[0112] In one embodiment of the present invention, a storage medium is provided, specifically a computer-readable storage medium (Memory). The computer-readable storage medium is a memory device in a computer device for storing programs and data. It is understandable that the computer-readable storage medium herein may include both built-in storage media in the computer device and, of course, extended storage media supported by the computer device. The computer-readable storage medium provides storage space, which stores the terminal's operating system. In addition, the storage space also stores one or more instructions suitable for being loaded and executed by a processor. These instructions may be one or more computer programs (including program code). It should be noted that the computer-readable storage medium herein may be a high-speed RAM (Random Access Memory) memory or a non-volatile memory, such as at least one disk storage device. The processor may load and execute one or more instructions stored in the computer-readable storage medium to implement the corresponding steps of the federated learning method for filtering noisy label samples in the above-mentioned embodiment.

[0113] Those skilled in the art will appreciate that the embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the present application can adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application can adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, optical storage, etc.) that contain computer-usable program code.

[0114] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the steps in the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0115] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0116] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0117] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents. Any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be covered by the scope of protection of the claims of the present invention.

Claims

1. A federated learning method capable of filtering noisy label samples, characterized in that: The following steps are involved: In each round of federated learning training, the selected client downloads the parameters of the global model from the server; For each selected client, the local personalized model and global model are trained using locally selected non-noise labeled samples. Upon completion of training, the parameters of the locally trained global model are updated and uploaded to the server. The personalized model is used to clean local data to select non-noise labeled sample data for local training. During training, the local personalized model and the global model are updated alternately by selecting samples. The server aggregates the parameter updates uploaded by each client and starts the next round of federated learning.

2. A federated learning method capable of filtering noisy label samples according to claim 1, characterized in that: In the step of using the locally selected non-noise label samples to train the local personalized model and the global model, the client's local personalized model and the global model are trained collaboratively, the two models are trained for the same number of iterations, and training and parameter updates are performed in each iteration.

3. The federated learning method capable of filtering noisy label samples according to claim 1, characterized in that: The step of using the locally selected non-noise label samples to train the local personalized model and the global model specifically includes: The sample selector is used to calculate and select non-noise label samples, and the loss function values of the client's local global model and personalized model on the non-noise label samples are calculated respectively. The parameters of the client's local global model and personalized model are calculated and updated.

4. A federated learning method capable of filtering noisy label samples according to claim 3, characterized in that: In the step of calculating and selecting non-noise label samples by the sample selector, The formula for selecting non-noise label samples is expressed as: Where, v n Indicates whether the example n is a noise label sample, v n =1, then example n is a non-noise label sample, v n = 0, then example n is a noise label sample; l CE (·) is the cross entropy loss function; L is the category of the label; x n is the characteristic of example n; is the label of example n; f(·) is the logistic regression classifier; The parameters of the logistic regression classifier f(·) are calculated using the following formula: Where, is a given function space; n k is the data sample selected by client k, is the mean number of data samples selected by client k; Where β is a predetermined hyperparameter; For the conditional distribution Lower your expectations, For a given noise dataset and noise labels The probability distribution of .

5. A federated learning method capable of filtering noisy label samples according to claim 4, characterized in that: In the step of respectively calculating the loss function values of the client's local global model and the personalized model on the non-noise label samples, First, initialize the client's local personalized model parameters. The initialized parameters are consistent with the parameters of the global model downloaded by the client from the server. The objective formula F(·) for global model training is as follows: Where w is the parameter of the global model; is the total number of samples selected on all clients; F k (·) represents the target formula during local client training, which is used to characterize the prediction loss of local data; K is the total number of clients; F k (·)The calculation formula is as follows: Where, v n Indicates whether the example n is a noise label sample, l(·) is the loss function, x n is the feature of example n, is the label of example n, and λ represents the parameter that controls the influence of the global model on the personalized model.

6. A federated learning method capable of filtering noisy label samples according to claim 5, characterized in that: In the step of calculating and updating the parameters of the client's local global model and personalized model, The parameter update method of the personalized model in the client is expressed as: The parameter update method of the global model in the client is expressed as: w k ←w k -el(w k -θ(w k )); Where θ k is the personalized model of client k; w k is the parameter weight of the model held by client k; ζ is the local learning rate; is the gradient operator; l k (·) is the loss function of client k; η is the global learning rate; θ is the global model.

7. The federated learning method capable of filtering noisy label samples according to claim 1, characterized in that: In the step of the server aggregating parameter updates uploaded by each client, the calculation expression of the aggregation parameter operation is: Where, represents the model parameter update uploaded by client k in round t+1, where t represents the round number; γ∈[0,1] is a hyperparameter. The larger γ is, the more the server relies on the parameter update uploaded by the client when aggregating the client uploaded parameters and updating the global model parameters; w t is the global model parameter of server t round, w t+1 are the global model parameters of the server in round t+1.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, the federated learning method capable of filtering noise label samples is implemented as claimed in any one of claims 1 to 7.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the federated learning method capable of filtering noisy label samples according to any one of claims 1 to 7 is implemented.

Images