Federal forgetting learning method and device, equipment and medium

Abstract

The invention discloses a federal forgetting learning method and device, equipment and a medium, and belongs to the technical field of information security. The method comprises the following steps: based on model parameters of clients participating in federated learning, generating a model parameter discrimination matrix of the clients; aggregating the model parameter discrimination matrixes of the plurality of remaining clients to obtain an aggregated parameter discrimination matrix; updating the parameters of the global model by comparing the aggregation parameter discrimination matrix with a model parameter discrimination matrix of the target client so as to forget the target client; and the plurality of remaining clients take the forgotten global model as an initial model to carry out federation post-training. According to the method, the number of global model parameters needing to be adjusted is reduced through the discrimination matrix of the model parameters, the overall performance of the global model is ensured through parameter scaling, most of the performance of the global model after forgetting on the remaining clients is reserved, and meanwhile it is ensured that memory of forgetting content is not aroused in the model recovery stage.

Description

Technical Field

[0001] The present invention belongs to the technical field of information security and relates to a federated unlearning method, device, equipment and medium. Background Art

[0002] Federated Unlearning can combine each client to complete the forgetting goal under the framework of Federated Learning (FL), expanding the application scope of machine unlearning and effectively ensuring privacy and data security. However, there are still many challenges in achieving an effective forgetting goal in this distributed machine learning scenario of federated learning. To address the challenge of the inability to directly obtain the dataset in the distributed machine learning scenario, many studies have started to explore federated unlearning technology, focusing on extending the traditional machine unlearning concept to the distributed machine learning scenario.

[0003] In the federated unlearning scenario, the most effective way is to remove the target client and then perform federated learning on the remaining clients, but this approach incurs a huge training cost. Therefore, the goal of federated unlearning is to design a more efficient solution than full retraining while maintaining the performance of the model. To address the above challenges, a series of methods have been proposed to achieve efficient federated unlearning. Some methods store information related to the client training phase on the central server and complete forgetting using this stored additional information when a forgetting request is initiated. These methods avoid the time overhead caused by retraining but increase the additional storage resource overhead due to the reliance on the saved information. To solve the problem of over-reliance on storage, other methods complete forgetting by optimizing the forgetting training process. For example, PGD achieves forgetting by reversing the training process of the model on the forgetting dataset, and FU_Active actively forgets the contributions of the target client through continuous learning. However, these methods usually introduce additional computational overhead and are difficult to meet in resource-constrained federated learning scenarios. Summary of the Invention

[0004] The purpose of the present invention is to overcome the deficiencies in the prior art and provide a federated unlearning method, device, equipment and medium that can achieve relatively thorough forgetting without excessive reliance on storage and computing resources.

[0005] To achieve the above object, the present invention is implemented by the following technical solutions:

[0006] In a first aspect, the present invention provides a federated unlearning method, including:

[0007] Generate a discriminant matrix of the model parameters for the clients participating in federated learning, where the clients include a target client and multiple remaining clients;

[0008] Aggregate the discriminant matrices of the model parameters of multiple remaining clients to obtain an aggregated parameter discriminant matrix;

[0009] Update the parameters of the global model by comparing the aggregated parameter discriminant matrix with the discriminant matrix of the model parameters of the target client to forget the target client;

[0010] Multiple remaining clients perform post-federated training with the global model that has completed forgetting as the initial model until the global model converges.

[0011] Further, the discriminant matrix of the model parameters includes the Fisher information matrix of the model parameters.

[0012] Further, the Fisher information matrix of the model parameters is expressed as:

[0013] I ( θ ) = E [ ∇ log p ( x | θ ) ( ∇ log p ( x | θ ) T )] ,

[0014] where, represents the Fisher information matrix of the model parameter , represents the probability distribution of the model parameter ; represents the gradient of the log-likelihood function of the model parameter ; E represents expectation; T represents matrix transpose.

[0015] Further, updating the parameters of the global model includes: scaling the parameters of the global model.

[0016] Further, scaling the parameters of the global model includes:

[0017] ,

[0018] ,

[0019] where, represents the -th parameter of the global model; represents the penalty coefficient of the parameter to be updated; represents the exponent of the penalty term; represents the weight of the parameter to be updated; represents the -th diagonal element of the discriminant matrix of the model parameters of the target client; represents the Diagonal elements.

[0020] Further, the multiple remaining clients perform federated post-training with the completed forgotten global model as the initial model, including:

[0021] In the first training round, send the completed forgotten global model to the multiple remaining clients to make it the initial local model for the first training round of the remaining clients;

[0022] The remaining clients separately train the initial local model using their local data to obtain local model parameters;

[0023] Aggregate the local model parameters of the multiple remaining clients to update the global model.

[0024] Further, the multiple remaining clients perform federated post-training with the completed forgotten global model as the initial model, and further include:

[0025] Send the global model updated in each training round to the multiple remaining clients to make it the initial local model for the next training round of the remaining clients, and perform the next round of training until the global model converges.

[0026] Further, aggregating the local model parameters of the multiple remaining clients to update the global model includes:

[0027] ,

[0028] Wherein, represents the local model parameters of the remaining client , represents the updated global model parameters, represents the number of remaining clients; represents the number of local datasets of the remaining client, represents the total data volume of the local datasets of the remaining clients.

[0029] In a second aspect, the present invention further provides a federated forgetting learning device, and the device includes:

[0030] A model parameter discrimination matrix generation module, configured to generate a model parameter discrimination matrix of a client based on the model parameters of the clients participating in federated learning, where the clients include a target client and multiple remaining clients;

[0031] An aggregated parameter discrimination matrix acquisition module, configured to aggregate the model parameter discrimination matrices of the multiple remaining clients to obtain an aggregated parameter discrimination matrix;

[0032] A target client forgetting module, which is used to update the parameters of the global model by comparing the aggregated parameter discrimination matrix with the model parameter discrimination matrix of the target client, so as to forget the target client;

[0033] A remaining client federated post-training module, which is used for multiple remaining clients to perform federated post-training with the global model after forgetting as the initial model until the global model converges.

[0034] In a third aspect, the present invention also provides a computer device, including:

[0035] A memory, which is used to store a computer program;

[0036] A processor, which is used to execute the computer program to implement the steps of the above-mentioned federated forgetting learning method.

[0037] In a fourth aspect, the present invention also provides a computer-readable storage medium, on which a computer program is stored, and the program is executed by a processor to implement the steps of the above-mentioned federated forgetting learning method

[0038] Compared with the prior art, the beneficial effects achieved by the present invention:

[0039] The federated forgetting learning method provided by the present invention reduces the number of global model parameters to be adjusted by using the Fisher information matrix of the model parameters as the discrimination matrix, thereby avoiding a significant impact on the global model, and ensures the overall performance of the global model by using parameter scaling. It not only effectively limits the range of parameter adjustment, but also retains most of the performance of the global model after forgetting on the remaining clients; during the model recovery process, using the global model after forgetting as the initial model for the remaining clients to perform federated post-training can greatly reduce the training cost required for federated post-training, and ensure that the global model will not evoke memories of the forgotten content during the recovery stage. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 It is a schematic flowchart of a federated forgetting learning method provided by an embodiment of the present invention;

[0041] Figure 2 It is a schematic framework diagram of the federated forgetting learning method in an embodiment of the present invention;

[0042] Figure 3 It is a schematic framework diagram of forgetting and post-training of the federated forgetting learning method in an embodiment of the present invention;

[0043] Figure 4 It is a schematic structural diagram of a federated forgetting learning device provided by an embodiment of the present invention;

[0044] Figure 5Internal structure diagram of the computer device provided by the embodiment of the present invention. Detailed implementation mode

[0045] The technical solution of the present invention will be described in detail below through the accompanying drawings and specific embodiments. The same reference numerals in the drawings denote the same or similar components or parts. Those skilled in the art should understand that these drawings are not necessarily drawn to scale. The embodiments of the present application and the specific features in the embodiments are detailed descriptions of the technical solution of the present application, rather than limitations on the technical solution of the present application. Without conflict, the technical features in the embodiments of the present application and the embodiments can be combined with each other.

[0046] The term "and / or" in this article is only a description of the association relationship of associated objects, indicating that there can be three relationships. For example, A and / or B can represent: A exists alone, A and B exist simultaneously, and B exists alone. In addition, the character " / " in this article generally represents an "or" relationship between the front and back associated objects.

[0047] Embodiment 1:

[0048] As Figures 1 to 3 shown, the embodiment of the present invention provides a federated unlearning method. Figure 1 is a schematic flow chart of the federated unlearning method. This flow chart only shows the logical order of the method described in this embodiment. On the premise of non-conflict, in other possible embodiments of the present invention, the steps shown or described can be completed in a different order from Figure 1 shown.

[0049] The federated unlearning method provided in this embodiment can be applied to a terminal and can be executed by a federated unlearning device. This device can be implemented in a software and / or hardware manner and can be integrated in the terminal.

[0050] See Figure 1 , the method of the embodiment of the present invention specifically includes the following steps:

[0051] Step 1: Generate a model parameter discrimination matrix for the client based on the model parameters of the clients participating in the federated learning. The clients include the target client and multiple remaining clients.

[0052] As Figure 2 shown, federated unlearning is usually carried out after federated learning is completed. In the embodiment of the present invention, the participants in the federated learning include a central server and There are multiple clients, each with its own independent local dataset, and the data between clients is not interoperable. After completing federated training, for privacy reasons, the target client sends a forgetting request to the server, asking to erase the contribution of its local dataset to the global model. The central server is responsible for receiving the forgetting requests from the clients. This server has low latency, high throughput, and high reliability, and can efficiently process and aggregate the parameters of the models from multiple clients. When the target client makes a forgetting request, the central server will use the federated forgetting learning method of the present invention to erase the influence of the target client's data on the global model.

[0053] The discriminant matrix of the model parameters of the present invention is the Fisher information matrix of the model parameters, which is used to determine whether a certain parameter is used to identify the data characteristics of a specific client.

[0054] After completing federated training, a global model is obtained, and this model has obtained the optimal parameters. These parameters are obtained by aggregating the parameters uploaded by multiple clients, and different model parameters are updated based on the data characteristics of each client to determine the classification results of specific client data characteristics. Therefore, the parameters aggregated by the target client can be identified and modified to eliminate the contribution of the target dataset to the global model. The Fisher Information Matrix (FIM) provides a method to measure the contribution of a specific dataset to the global model. It uses the first-order partial derivatives generated during the training process to calculate the importance, greatly reducing the computational overhead.

[0055] The Fisher information matrix of the model parameters is expressed as:

[0056] I ( θ ) = E [ ∇ log p ( x | θ ) ( ∇ log p ( x | θ ) T )] ,

[0057] where, represents the Fisher information matrix of the model parameter , represents the probability distribution of the model parameter ; represents the log-likelihood function of the model parameter ; represents the gradient; E represents the expectation; T represents the matrix transpose.

[0058] Step 2: Aggregate the discriminant matrices of the model parameters of multiple remaining clients to obtain an aggregated parameter discriminant matrix.

[0059] After the central server collects the discriminant matrices of the model parameters of multiple remaining clients, it performs weighted aggregation according to the sizes of the local datasets of each remaining client to obtain an aggregated parameter discriminant matrix.

[0060] Step 3: Update the parameters of the global model by comparing the aggregated parameter discrimination matrix with the model parameter discrimination matrix of the target client to forget the target client.

[0061] The commonly used method in the prior art to update the global model parameters is:

[0062] ,

[0063] where, represents the th parameter of the global model; represents the th diagonal element of the model parameter discrimination matrix of the target client.

[0064] The above parameter update method realizes forgetting by directly setting the parameters with importance greater than zero to zero. This adjustment method is the most intuitive. However, when there are the same categories in the datasets of the target client and the remaining clients, the judgment criterion based on comparison with zero will cause too many parameters to be adjusted, and directly adjusting the parameters to zero will significantly affect the model performance. The challenge of federated forgetting learning lies in how to maintain the performance of the model on the remaining clients during the forgetting process. Therefore, the present invention adopts a more stringent judgment criterion and a more detailed parameter adjustment method.

[0065] The present invention realizes the update of the parameters of the global model by scaling the parameters of the global model.

[0066] Specifically, the way to scale the parameters of the global model in the present invention is:

[0067] ,

[0068] ,

[0069] where, represents the th parameter of the global model; represents the penalty coefficient of the parameter to be updated; represents the exponent of the penalty term; represents the weight of the parameter to be updated; represents the th diagonal element of the model parameter discrimination matrix of the target client; represents the th diagonal element of the aggregated parameter discrimination matrix.

[0070] Step 4: Multiple remaining clients perform federated post-training with the global model that has completed forgetting as the initial model until the global model converges.

[0071] Although the performance of the global model has been deliberately maintained during the forgetting process, inevitably, the forgetting process still affects the evaluation of the global model on the remaining dataset. Therefore, after the forgetting process is completed, the global server often initiates model recovery, that is, performing a few rounds of standard federated learning on the remaining clients.

[0072] During the forgetting process, the present invention uses the Fisher information matrix as the discrimination matrix to reduce the number of parameters to be adjusted, thereby avoiding a significant impact on the global model, and uses parameter scaling to ensure the overall performance of the global model. It not only effectively limits the range of parameter adjustment but also retains most of the performance of the model on the remaining clients. Therefore, during the model recovery process, using the forgotten global model as the initial model for post-federated training can greatly reduce the training cost during the recovery process.

[0073] The remaining clients use the forgotten global model as the initial model for post-federated training, specifically including: in each training round, sending the global model to the remaining clients, making it the local model of the remaining clients; the remaining clients use their local data to train the local model separately to obtain local model parameters; aggregating the local model parameters of multiple remaining clients to update the global model. Sending the updated global model to the remaining clients, making it the local model of the remaining clients in the next round of training, and performing the next round of training until the global model converges.

[0074] In the first training round, send the forgotten global model to the remaining clients, making it the initial local model of the remaining clients in the first training round; each remaining client uses its local data to train the initial local model separately to obtain local model parameters; aggregating the local model parameters of the remaining clients to perform the first update of the global model.

[0075] Thereafter, in each training round, send the updated global model to the remaining clients, making it the initial local model of the remaining clients in the next training round, and performing the next round of training until the global model converges or reaches the preset number of training rounds.

[0076] Specifically, aggregating the local model parameters of multiple remaining clients to update the global model includes:

[0077] ,

[0078] wherein, represents the remaining clients local model parameters represent the updated global model parameters represents the number of remaining clients; represents the remaining clients the number of local datasets of represents the total data volume of the local datasets of

[0079] Figure 3 Fig. shows a framework diagram of forgetting and post-training for the federated forgetting learning method provided by an embodiment of the present invention.

[0080] Embodiments of the present invention designed a simulation experiment of the federated forgetting learning method in a real scenario. The experimental datasets used were the MNIST dataset and the CIFAR10 dataset. The network model used consisted of six convolutional layers, three max pooling layers, a fully connected layer with a ReLU activation function, and a final softmax output layer. To evaluate the forgetting effect of the model, the backdoor attack success rate was used as an evaluation metric (B_A). At the same time, to test the performance of the model on the remaining clients, the clean dataset accuracy (C_A) was also used as another evaluation metric. After forgetting, the lower the backdoor attack success rate and the higher the clean dataset accuracy, the better the forgetting effect.

[0081] The federated forgetting learning method of the present invention was compared with three existing federated forgetting learning methods, namely FedAvg, Federated Unlearning Retrain (FUR), and Projected Gradient Descent (PGD). Among them, FedAvg was used as the original model without forgetting operation for comparison with other methods. FUR represents retraining from scratch to obtain a clean model without the information of the target client; PGD uses the method of gradient ascent for forgetting to obtain a forgetting model.

[0082] The Dirichlet distribution method based on label imbalance was used to allocate the dataset to each client. When the number of clients was 10, at the Dirichlet distribution concentration parameters , , the backdoor accuracy and the clean dataset accuracy of each algorithm are shown in Table 1. Among them, the larger the concentration parameter value, the lower the data heterogeneity. It can be found that the model of the present invention achieves effective forgetting while maintaining the performance of the model after forgetting on the remaining clients. For example, on the CIFAR-10 dataset, the FUR method at When B_A was 10.95%, the method of the present invention was 13.19%, and PGD was 23.39%. The reason for this phenomenon is that the method of the present invention compares and instead of and 0, and through the adjustment based on the scaling rule, rather than directly changing to zero, thus achieving experimental results similar to those of FUR.

[0083] Table 1 Backdoor accuracy (B_A) and clean accuracy (C_A) of different forgetting methods on different datasets

[0084]

[0085] To alleviate the impact on the model during the forgetting process, model restoration is required, and the number of training rounds required for restoration by the method of the present invention is far better than the comparative method. The specific convergence rounds are shown in Table 2. Although the PGD method has an improvement compared to the FUR retraining, the training cost is still relatively high. For example, the number of rounds required for convergence of the present method on the CIFAR-10 dataset is 3.5 times higher than that of the FUR method, while the PGD method is only about 1.5 times. Therefore, the method of the present invention can effectively reduce the restoration cost.

[0086] Table 2 Training rounds of different forgetting methods

[0087]

[0088] Example 2:

[0089] Based on the same inventive concept as Example 1, the embodiment of the present invention also provides a federated forgetting learning device for implementing the above-mentioned federated forgetting learning method. The implementation solution provided by this device to solve problems is similar to the implementation solution described in the above method. Therefore, the specific limitations in the embodiment of the federated forgetting learning device provided below can refer to the limitations on the federated forgetting learning method in the above text, and will not be repeated here.

[0090] As Figure 4 shown, the embodiment of the present invention provides a federated forgetting learning device, including:

[0091] A model parameter discrimination matrix generation module, configured to generate a model parameter discrimination matrix of a client based on the model parameters of the clients participating in the federated learning, where the clients include a target client and multiple remaining clients;

[0092] An aggregated parameter discrimination matrix acquisition module, configured to aggregate the model parameter discrimination matrices of multiple remaining clients to obtain an aggregated parameter discrimination matrix;

[0093] A target client forgetting module, which is used to update the parameters of the global model by comparing the aggregated parameter discrimination matrix with the model parameter discrimination matrix of the target client, so as to forget the target client;

[0094] A remaining client federated post-training module, which is used for multiple remaining clients to perform federated post-training with the global model that has completed forgetting as the initial model until the global model converges.

[0095] Embodiment 3:

[0096] The embodiment of the present invention also provides a computer device, which can be a server, and its internal structure diagram can be as Figure 5 shown. The computer device includes a processor, a memory, an input / output interface (Input / Output, abbreviated as I / O), and a communication interface. Among them, the processor, the memory, and the input / output interface are connected through a system bus, and the communication interface is connected to the system bus through the input / output interface. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program, and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The input / output interface of the computer device is used for exchanging information between the processor and external devices. The communication interface of the computer device is used for communicating with external terminals through a network connection. When the computer program is executed by the processor, it realizes the federated forgetting learning method in the foregoing embodiments.

[0097] Those skilled in the art can understand that Figure 5 the structure shown in is only a block diagram of some structures related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine some components, or have different component arrangements.

[0098] Embodiment 4:

[0099] The embodiment of the present invention also provides a computer-readable storage medium, on which a computer program is stored, and when the program is executed by a processor, it realizes the steps of the following method:

[0100] Generate a model parameter discrimination matrix of a client based on the model parameters of the clients participating in federated learning, where the clients include a target client and multiple remaining clients;

[0101] Aggregate the model parameter discrimination matrices of multiple remaining clients to obtain an aggregated parameter discrimination matrix;

[0102] The parameters of the global model are updated by comparing the aggregation parameter discrimination matrix with the model parameter discrimination matrix of the target client, so as to forget the target client.

[0103] The multiple remaining clients perform federated post-training with the global model that has completed forgetting as the initial model until the global model converges.

[0104] Those skilled in the art should understand that the embodiments of the present invention can be provided as a method, a system, or a computer program product. Therefore, the present invention can adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present invention can adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk memory, CD-ROM, optical memory, etc.) containing computer-usable program code.

[0105] The present invention is described with reference to the flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to the embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing devices generate means for implementing the functions specified in one process Figure 1 one process or multiple processes and / or blocks Figure 1 one block or multiple blocks.

[0106] These computer program instructions can 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 generate a manufactured product including instruction means, and the instruction means implements the functions specified in one process Figure 1 one process or multiple processes and / or blocks Figure 1 one block or multiple blocks.

[0107] These computer program instructions can also be loaded onto a computer or other programmable data processing device, so that a series of operation steps are executed on the computer or other programmable device to generate a computer-implemented process. Thus, the instructions executed on the computer or other programmable device provide steps for implementing the functions specified in one process Figure 1 one process or multiple processes and / or blocks Figure 1 one block or multiple blocks.

[0108] The embodiments of the present invention have been described above in conjunction with the accompanying drawings. However, the present invention is not limited to the above specific embodiments. The above specific embodiments are merely illustrative and not restrictive. Under the inspiration of the present invention, those of ordinary skill in the art can also make many forms without departing from the spirit of the present invention and the scope protected by the claims. All of these are within the protection scope of the present invention.

Claims

1. A federated forgetting learning method, characterized in that, Including: Generating a model parameter discrimination matrix for a client based on the model parameters of the clients participating in federated learning, where the clients include a target client and multiple remaining clients; Aggregating the model parameter discrimination matrices of multiple remaining clients to obtain an aggregated parameter discrimination matrix; Updating the parameters of the global model by comparing the aggregated parameter discrimination matrix with the model parameter discrimination matrix of the target client to forget the target client; Performing post-federation training on multiple remaining clients with the global model after forgetting as the initial model until the global model converges.

2. The federated forgetting learning method according to claim 1, wherein The model parameter discrimination matrix includes the Fisher information matrix of the model parameters.

3. The federated forgetting learning method according to claim 2, wherein The Fisher information matrix of the model parameters is expressed as: ， wherein, represents the Fisher information matrix of the model parameters; represents the probability distribution of the model parameters; represents the gradient of the log-likelihood function of the model parameters; E represents expectation; T represents matrix transpose.

4. The federated forgetting learning method according to claim 2, wherein, Updating the parameters of the global model includes: scaling the parameters of the global model, including: ， ， Among them, represents the th parameter of the global model; represents the penalty coefficient of the parameter to be updated; represents the exponent of the penalty term; represents the weight of the parameter to be updated; represents the th diagonal element of the model parameter discrimination matrix of the target client; represents the th diagonal element of the aggregated parameter discrimination matrix.

5. The federated forgetting learning method according to claim 1, wherein Performing post-federation training on multiple remaining clients with the global model after forgetting as the initial model includes: In the first training round, sending the global model after forgetting to multiple remaining clients to make it the initial local model of the remaining clients; The remaining clients separately train the initial local model with their local data to obtain local model parameters; Aggregating the local model parameters of multiple remaining clients to update the global model.

6. The federated forgetting learning method according to claim 5, wherein Performing post-federation training on multiple remaining clients with the global model after forgetting as the initial model further includes: Sending the global model updated in each training round to multiple remaining clients to make it the initial local model for the next training round of the remaining clients for the next round of training until the global model converges.

7. The federated forgetting learning method according to claim 5, wherein Aggregating the local model parameters of multiple remaining clients to update the global model includes: ， Among them, represents the local model parameters of the remaining clients , represents the updated global model parameters, represents the number of remaining clients; represents the remaining clients 's local dataset quantity, represents the total data volume of the local datasets of the remaining clients.

8. A federated forgetting learning device, characterized in that, Including: A model parameter discrimination matrix generation module for generating a model parameter discrimination matrix for a client based on the model parameters of the clients participating in federated learning, where the clients include a target client and multiple remaining clients; An aggregated parameter discrimination matrix acquisition module for aggregating the model parameter discrimination matrices of multiple remaining clients to obtain an aggregated parameter discrimination matrix; A target client forgetting module for updating the parameters of the global model by comparing the aggregated parameter discrimination matrix with the model parameter discrimination matrix of the target client to forget the target client; A remaining client post-federation training module for performing post-federation training on multiple remaining clients with the global model after forgetting as the initial model until the global model converges.

9. A computer device, characterized in that, Including: A memory for storing a computer program; A processor for executing the computer program to implement the steps of the federated forgetting learning method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps of the federated forgetting learning method according to any one of claims 1 to 7.

Images