Federal forgetting learning method and system based on prototype

Abstract

The invention discloses a federal forgetting learning method and system based on a prototype, and belongs to the technical field of federal forgetting learning.The method comprises the steps that firstly, a forgetting strategy based on the prototype is provided, and the forgetting process is guided through the prototype of a target client; the data forgetting is realized by maximizing the prototype loss between the remaining clients and the target client; and a model recovery method based on a prototype is designed, and sample features in a feature space are adjusted by optimizing a global prototype, so that the accuracy of the model on the residual data set is quickly and effectively recovered. According to the method, on the premise that the global model performance is not affected, the forgetting requirement of the client side is met, and data privacy is effectively guaranteed.

Description

Technical Field

[0001] The present invention relates to a prototype-based federated forgetting learning method and system, belonging to the technical field of federated forgetting learning technology. Background Art

[0002] Federated Learning (FL) is a promising distributed machine learning approach that allows multiple participants to collaboratively train a global model without sharing data. This approach not only efficiently utilizes distributed computing resources but also effectively protects user privacy, making it suitable for a variety of application scenarios. However, in FL, situations such as client disconnection, data errors, and protocol expiration may require "forgetting" specific datasets. To this end, many machine learning forgetting algorithms have been proposed to achieve effective forgetting of specific data. Some studies directly eliminate the impact of deleted data through reverse gradient operations, but these methods are generally designed for centralized environments. In a distributed environment, data is distributed across various clients, and traditional centralized forgetting algorithms have difficulty managing and executing data forgetting efficiently. To address this challenge, researchers began to extend traditional machine forgetting to distributed scenarios, thereby developing federated forgetting technology.

[0003] In federated forgetting, the most straightforward and effective approach is to retrain the remaining clients from scratch, but this approach incurs significant computational, communication, and time overhead. Therefore, the goal of federated forgetting is to develop a solution that is more efficient than retraining while maintaining model effectiveness. To this end, recent research has focused on two approaches: parameter modification and training modification. Parameter modification methods remove the influence of a specific dataset on the global model by directly deleting historical gradient updates. Model parameters are then adjusted through partial or complete retraining to compensate for performance losses caused by forgetting. In contrast, training modification methods do not directly modify model parameters, but instead implement data forgetting through a gradual smoothing process. This can be achieved by adjusting the optimization objective, modifying the loss function, or introducing constraints or regularization terms when forgetting data to protect or adjust model parameters. For example, gradient ascent can be used to optimize the loss function to maximize a specific objective to promote forgetting. However, these federated forgetting methods typically require significant storage resources or additional data resources, which remains challenging in practical applications. Summary of the Invention

[0004] The purpose of the present invention is to provide a prototype-based federated forgetting learning method and system, which guides the forgetting process through the prototype of the target client and effectively achieves data forgetting by maximizing the prototype loss between the remaining clients and the target client; and optimizes the sample features learned by the global prototype model in the feature space to quickly and effectively restore the model accuracy on the remaining data set, ensuring that the client's forgetting needs are met without affecting the performance of the global model.

[0005] In order to achieve the above objectives / solve the above technical problems, the present invention is implemented by adopting the following technical solutions.

[0006] In one aspect, the present invention provides a prototype-based federated forgetting learning method, comprising:

[0007] Obtain the local prototype and forget request calculated as the target prototype when the target client is ready to exit the initial federated learning training phase;

[0008] Based on the forget request, the aggregation process of the federated learning training phase is interrupted and the forget phase is entered;

[0009] In the forgetting phase, the remaining clients adopt a target prototype-based model forgetting strategy to eliminate the impact of the target client data on the local models of the remaining clients, and introduce a global prototype to perform model recovery training on the local prototype after the impact is eliminated;

[0010] The local models of the remaining clients after the resumed training are obtained and aggregated to obtain the reconstructed global model, and the initial federated learning training phase is resumed based on the reconstructed global model.

[0011] Furthermore, the remaining clients adopt a model forgetting strategy based on the target prototype, specifically including:

[0012] The local forgetting loss function is constructed by the target prototype. The specific expression is:

[0013] ;

[0014] in: is the forgetting loss function of the local model of the i-th remaining client in the local dataset, is the local dataset of the i-th remaining client, is the local model of the i-th remaining client; is the prototype loss function between the local prototype and the target prototype, is the local prototype of the i-th remaining client, For the target prototype; is the standard cross entropy loss function of the local model of the i-th remaining client, is the local model of the i-th remaining client For input samples The predicted output of , is the input sample The corresponding sample labels,

[0015] The local models of the remaining clients are guided to forget the target client data by maximizing the local forgetting loss function, and the local models of the remaining clients that have forgotten the target client data are updated by performing a gradient ascent step. The expression is:

[0016] ;

[0017] ;

[0018] in: The target client Class target prototype, Use the local model for the i-th remaining client For input samples The embedding representation of is: is a candidate local model in the remaining client local model space, is the space where the remaining client candidate local models are located, is the dimension of the remaining client candidate local model parameters, is the set hyperparameter, is the updated local model of the i-th remaining client, is the set learning rate, is the gradient of the local loss function with respect to the local model of the i-th remaining client on the local dataset;

[0019] This strategy effectively forgets sensitive information while protecting data privacy by introducing prototypes as alignment references rather than directly using sensitive data samples. In addition, the prototype constraint provides a clear direction for the optimization process, which can more efficiently guide the model away from the data feature areas of the target client and accelerate the forgetting process.

[0020] Furthermore, the expression of the gradient of the local loss function with respect to the local model of the i-th remaining client on the local dataset is:

[0021] .

[0022] Furthermore, the target prototype is specifically expressed as:

[0023] ;

[0024] in, Target Client The local data set belongs to A subset of the class, Input sample for the local model of target client j The embedding representation of is the local model of target client j, and K is the total number of prototype categories.

[0025] Furthermore, the prototype loss function between the local prototype and the target prototype is The specific calculation method is:

[0026] , ;

[0027] in: Input sample for the local model of the i-th remaining client The target prototype loss function constrains the sample embedding representation extracted by the client model to be far away from the target prototype, thereby suppressing the model's ability to express target features in the embedding space, thereby achieving forgetting at the semantic level.

[0028] Furthermore, the introduction of the global prototype to perform model recovery training on the local prototype after eliminating the influence specifically includes:

[0029] The global prototype is introduced to construct the recovery loss function, which is expressed as:

[0030] ;

[0031] ;

[0032] in: is the recovery loss function of the updated local model of the i-th remaining client on the local dataset, is the cross entropy loss function of the local model standard of the i-th remaining client after update, For the The remaining clients forget the local model after training and updating the input sample The predicted output is the input sample The corresponding sample labels, is the prototype loss function between the local prototype and the global prototype of the i-th remaining client, is the global prototype, is the set weight coefficient, For the Class global prototype, Indicates that the The client collection of the class global prototype, For the Of the remaining clients Class local prototype, K is the total number of prototype categories;

[0033] According to the recovery loss function, the performance of the updated remaining client local model is optimized in combination with the gradient descent method. The expression is:

[0034] ;

[0035] in: For the The remaining clients perform performance optimization on the local model. is the gradient of the recovery loss function of the updated local model of the i-th remaining client on the local dataset;

[0036] By aligning local prototypes with the aggregated global prototype, the recovery process effectively alleviates the disturbance of model structure and feature expression caused by forgetting, helping the model to recapture stable and shared semantic representations; the prototype-guided approach has directionality and generalization capabilities, allowing the model to achieve performance recovery within a small number of iterations, effectively avoiding catastrophic forgetting.

[0037] Furthermore, the gradient of the recovery loss function of the updated local model of the i-th remaining client on the local dataset is specifically expressed as:

[0038] .

[0039] In a second aspect, the present invention provides a prototype-based federated forgetting learning device, comprising:

[0040] An acquisition module is used to obtain the local prototype and forget request calculated as the target prototype when the target client is ready to exit the initial federated learning training phase;

[0041] Interrupt module: used to interrupt the aggregation process of the federated learning training phase and enter the forgetting phase based on the forgetting request;

[0042] A processing module is used to, during the forgetting phase, enable the remaining clients to adopt a model forgetting strategy based on the target prototype to eliminate the influence of the target client data on the local models of the remaining clients, and introduce a global prototype to perform model recovery training on the local prototype after the influence is eliminated;

[0043] The recovery module is used to obtain the local models of the remaining clients after the recovery training and aggregate them to obtain the reconstructed global model, and restore the initial federated learning training phase based on the reconstructed global model.

[0044] In a third aspect, the present invention provides a prototype-based federated forgetting learning system, comprising:

[0045] Memory, used to store computer programs / instructions;

[0046] A processor is used to execute the computer program / instructions to implement the steps of the above-mentioned prototype-based federated forgetting learning method.

[0047] In a fourth aspect, the present invention provides a computer-readable storage medium having a computer program / instruction stored thereon, characterized in that when the computer program / instruction is executed by a processor, the steps of the above-mentioned prototype-based federated forgetting learning method are implemented.

[0048] Compared with existing technologies, the present invention achieves the following beneficial effects: the present invention proposes to use the prototype of the target client to guide the forgetting process and achieves the forgetting of specific data by maximizing the prototype loss between the remaining clients and the deleted clients. The method of the present invention does not require the storage of additional historical parameters or gradient updates, thereby decoupling the forgetting process from FL training and achieving efficient federated forgetting learning.

[0049] The present invention also implements a concise model recovery training method that focuses on exchanging prototypes rather than model updates and uses global prototypes for guidance, reducing communication overhead, accelerating the recovery process, and improving the performance of the model after forgetting. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 This is a schematic diagram of the system framework of a prototype-based federated forgetting learning method designed by the present invention;

[0051] Figure 2 is a schematic diagram of the backdoor dataset accuracy of each algorithm on the MNIST dataset under different numbers of clients in this embodiment;

[0052] Figure 3 is a schematic diagram of the clean dataset accuracy of each algorithm on the MNIST dataset with different numbers of clients in this embodiment. DETAILED DESCRIPTION

[0053] It should be noted that:

[0054] The technical solution of the present invention is described in detail below through the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations on the technical solution of the present invention. In the absence of conflict, the embodiments of the present invention and the technical features in the embodiments can be combined with each other.

[0055] The term "and / or" simply describes a relationship between related objects, indicating that three possible relationships exist. For example, "A and / or B" can mean: A exists alone, A and B exist simultaneously, or B exists alone. Additionally, the character " / " generally indicates an "or" relationship between the related objects.

[0056] Example 1

[0057] like Figure 1 In the embodiment 1 shown, this embodiment provides a prototype-based federated forgetting learning method, including:

[0058] When the target client is ready to exit the initial federated learning training phase, it calculates the local prototype and uploads the calculated local prototype as the target prototype together with the forget request to the central server;

[0059] After receiving the forget request from the target client, the central server interrupts the aggregation process of the federated learning training phase and enters the forget phase;

[0060] In the forgetting phase, the remaining clients adopt a model forgetting strategy based on the target prototype to eliminate the influence of the target client data on the local models of the remaining clients;

[0061] The remaining clients introduce the global prototype to perform model recovery training on the local prototype after eliminating the impact;

[0062] The central server collects the local models of the remaining clients after the training is resumed and aggregates them to obtain the reconstructed global model. The central server resumes the initial federated learning training phase based on the reconstructed global model.

[0063] The remaining clients adopt a model forgetting strategy based on the target prototype, specifically including:

[0064] The local forgetting loss function is constructed by the target prototype. The specific expression is:

[0065] ;

[0066] in: is the forgetting loss function of the local model of the i-th remaining client in the local dataset, is the local dataset of the i-th remaining client, is the local model of the i-th remaining client; is the prototype loss function between the local prototype and the target prototype, is the local prototype of the i-th remaining client, For the target prototype; is the standard cross entropy loss function of the local model of the i-th remaining client, is the local model of the i-th remaining client For input samples The predicted output of , is the input sample The corresponding sample labels,

[0067] The local models of the remaining clients are guided to forget the target client data by maximizing the local forgetting loss function, and the local models of the remaining clients that have forgotten the target client data are updated by performing a gradient ascent step. The expression is:

[0068] ;

[0069] ;

[0070] in: The target client Class target prototype, Use the local model for the i-th remaining client For input samples The embedding representation of is: is a candidate local model in the remaining client local model space, is the space where the remaining client candidate local models are located, is the dimension of the remaining client candidate local model parameters, is the set hyperparameter, is the updated local model of the i-th remaining client, is the set learning rate, is the gradient of the local loss function with respect to the local model of the i-th remaining client on the local dataset;

[0071] The expression of the gradient of the local loss function with respect to the local model of the i-th remaining client on the local dataset is:

[0072] .

[0073] The specific expression of the target prototype is:

[0074] ;

[0075] in, Target Client The local data set belongs to A subset of the class, Input sample for the local model of target client j The embedding representation of is the local model of target client j, and K is the total number of prototype categories.

[0076] The prototype loss function between the local prototype and the target prototype The specific calculation method is:

[0077] , ;

[0078] in: Input sample for the local model of the i-th remaining client Embedding representation of .

[0079] The remaining clients introduce the global prototype to perform model recovery training on the local prototype after eliminating the influence, specifically including:

[0080] The global prototype is introduced to construct the recovery loss function, which is expressed as:

[0081] ;

[0082] ;

[0083] in: is the recovery loss function of the updated local model of the i-th remaining client on the local dataset, is the cross entropy loss function of the local model standard of the i-th remaining client after update, For the The remaining clients forget the local model after training and updating the input sample The predicted output of is the prototype loss function between the local prototype and the global prototype of the i-th remaining client, is the global prototype, is the set weight coefficient, For the Class global prototype, Indicates that the The client collection of the class global prototype, For the Of the remaining clients Class local prototype;

[0084] According to the recovery loss function, the performance of the updated remaining client local model is optimized in combination with the gradient descent method. The expression is:

[0085] ;

[0086] in: For the The remaining clients perform performance optimization on the local model. is the gradient of the recovery loss function of the updated local model of the i-th remaining client on the local dataset.

[0087] The gradient of the recovery loss function of the updated local model of the i-th remaining client on the local dataset is specifically expressed as:

[0088] .

[0089] like Figures 2 and 3 As shown in the figure, based on the above content, the present invention designs a prototype-based federated forgetting learning method and conducts simulation experiments in actual scenarios; the two experimental datasets used are MNIST and CIFAR10 datasets, and for the MNIST dataset, a network consisting of 2 convolutional layers and 2 fully connected layers is used; for CIFAR10, the ResNet-18 model is used.

[0090] The experiment compared the proposed method with three existing federated learning algorithms: FedAvg, Federated Retrain (FR), and Projected Gradient Descent (PGD). FedAvg served as the original model, without any forgetting. FR, retrained from scratch to produce an ideal clean model, was often used as a benchmark for comparing forgetting effects. To assess the model's forgetting effect on target client data, the backdoor attack success rate was used as an evaluation metric. After using this forgetting method, a low backdoor attack success rate indicates that the target client's data information has been forgotten in the model.

[0091] The experimental results are detailed as follows: Based on the above invention, a prototype-based federated learning forgetting learning method is designed for practical application. When the target client initiates a forget request when the communication rounds reach 60% of the total rounds, the system forgets its historical contributions. Under experimental settings with different numbers of clients, the classification accuracy of each method on a clean test dataset and a backdoor-labeled test dataset is evaluated. The experimental results show that as the number of clients increases, the accuracy of each method on both test sets shows a downward trend. This phenomenon is mainly due to the increased degree of non-independent and identically distributed data brought about by the increase in the number of clients, which in turn weakens the generalization ability of the global model. It is worth noting that the method in this paper always maintains similar performance to the traditional FR method under different numbers of clients, fully demonstrating its stability and scalability in data heterogeneous environments.

[0092] When the number of clients is 10, the target client sends a forget request when the communication rounds reach 60% of the total rounds. The accuracy of the learning model of each method on the clean test dataset and the backdoor marked test dataset is shown in Table 1 below:

[0093] Table 1 Accuracy results of each algorithm on different test datasets on MNIST and CIFAR10 datasets

[0094] ;

[0095] Experimental results show that our method can effectively eliminate the influence of the target client. For example, on the MNIST dataset, the accuracy of the FedAvg model affected by the backdoor attack data is 72.80%. In comparison, the accuracy of the FR model is 11.42%, the method of our method is 12.43%, and the PGD model is 16.35%. Their forgetting rates are 100%, 98.35%, and 91.96%, respectively. Taking 98.35% as an example, the specific calculation of the forgetting rate is: (0.7280-0.1243) ÷ (0.7280-0.1142) = 0.9835.

[0096] Compared to PGD, the proposed method demonstrates lower backdoor attack accuracy, more complete forgetting, and better model accuracy. On the MNIST dataset, the proposed method achieves nearly identical backdoor attack accuracy to retraining and surpasses retraining on clean datasets. This phenomenon demonstrates that the proposed method, through prototype-driven model recovery, promotes further model convergence after forgetting the target client data, thereby improving overall performance and generalization.

[0097] Example 2

[0098] This embodiment provides a prototype-based federated forgetting learning device, including:

[0099] An acquisition module is used to obtain the local prototype and forget request calculated as the target prototype when the target client is ready to exit the initial federated learning training phase;

[0100] Interrupt module: used to interrupt the aggregation process of the federated learning training phase and enter the forgetting phase based on the forgetting request;

[0101] A processing module is used to, during the forgetting phase, enable the remaining clients to adopt a model forgetting strategy based on the target prototype to eliminate the influence of the target client data on the local models of the remaining clients, and introduce a global prototype to perform model recovery training on the local prototype after the influence is eliminated;

[0102] The recovery module is used to obtain the local models of the remaining clients after the recovery training and aggregate them to obtain the reconstructed global model, and restore the initial federated learning training phase based on the reconstructed global model.

[0103] Example 3

[0104] This embodiment provides a prototype-based federated forgetting learning system, including:

[0105] Memory, used to store computer programs / instructions;

[0106] A processor is used to execute the computer program / instructions to implement the steps of the prototype-based federated forgetting learning method described in Example 1.

[0107] Example 4

[0108] This embodiment provides a computer-readable storage medium having a computer program / instruction stored thereon, characterized in that when the computer program / instruction is executed by a processor, the steps of the prototype-based federated forgetting learning method described in Example 1 are implemented.

[0109] The present invention proposes to use the prototype of the target client to guide the forgetting process and achieve forgetting of specific data by maximizing the prototype loss between the remaining clients and the deleted clients. The method of the present invention does not require the storage of additional historical parameters or gradient updates, thereby decoupling the forgetting process from FL training and achieving efficient federated forgetting learning.

[0110] The present invention also implements a concise model recovery training method that focuses on exchanging prototypes rather than model updates and uses global prototypes for guidance, reducing communication overhead, accelerating the recovery process, and improving the performance of the model after forgetting.

[0111] The embodiments of the present invention are described above in conjunction with the accompanying drawings, but the present invention is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Under the guidance of the present invention, ordinary technicians in this field can also make many forms without departing from the scope of protection of the purpose of the present invention and the claims, which are all protected by the present invention.

Claims

1. A prototype-based federated forgetting learning method, characterized in that: include: Obtain the local prototype and forget request calculated as the target prototype when the target client is ready to exit the initial federated learning training phase; Based on the forget request, the aggregation process of the federated learning training phase is interrupted and the forget phase is entered; In the forgetting phase, the remaining clients adopt a target prototype-based model forgetting strategy to eliminate the impact of the target client data on the local models of the remaining clients, and introduce a global prototype to perform model recovery training on the local prototype after the impact is eliminated; The local models of the remaining clients after the resumed training are obtained and aggregated to obtain the reconstructed global model, and the initial federated learning training phase is resumed based on the reconstructed global model.

2. The prototype-based federated forgetting learning method according to claim 1, characterized in that: The remaining clients adopt a model forgetting strategy based on the target prototype, specifically including: The local forgetting loss function is constructed by the target prototype. The specific expression is: ； in: is the forgetting loss function of the local model of the i-th remaining client in the local dataset, is the local dataset of the i-th remaining client, is the local model of the i-th remaining client; is the prototype loss function between the local prototype and the target prototype, is the local prototype of the i-th remaining client, For the target prototype; is the standard cross entropy loss function of the local model of the i-th remaining client, is the local model of the i-th remaining client For input samples The predicted output of , is the input sample Corresponding sample labels; The local models of the remaining clients are guided to forget the target client data by maximizing the local forgetting loss function, and the local models of the remaining clients that have forgotten the target client data are updated by performing a gradient ascent step. The expression is: ； ； in: The target client Class target prototype, Use the local model for the i-th remaining client For input samples The embedding representation of is: is a candidate local model in the remaining client local model space, is the space where the remaining client candidate local models are located, is the dimension of the remaining client candidate local model parameters, is the set hyperparameter, is the updated local model of the i-th remaining client, is the set learning rate, is the gradient of the local loss function with respect to the local model of the i-th remaining client on the local dataset.

3. The prototype-based federated forgetting learning method according to claim 2, characterized in that: The expression of the gradient of the local loss function with respect to the local model of the i-th remaining client on the local dataset is: 。 4. The prototype-based federated forgetting learning method according to claim 2, characterized in that: The specific expression of the target prototype is: ； in, Target Client The local data set belongs to A subset of the class, Input sample for the local model of target client j The embedding representation of is the local model of target client j, and K is the total number of prototype categories.

5. The prototype-based federated forgetting learning method according to claim 2, characterized in that: The prototype loss function between the local prototype and the target prototype The specific calculation method is: , ； in: Input sample for the local model of the i-th remaining client Embedding representation of .

6. The prototype-based federated forgetting learning method according to claim 1, characterized in that: The introduction of the global prototype to perform model recovery training on the local prototype after eliminating the influence specifically includes: The global prototype is introduced to construct the recovery loss function, which is expressed as: ； ； in: is the recovery loss function of the updated local model of the i-th remaining client on the local dataset, is the local dataset of the i-th remaining client, is the updated local model of the i-th remaining client, is the cross entropy loss function of the local model standard of the i-th remaining client after update, For the The remaining clients forget the local model after training and updating the input sample The predicted output of is the input sample The corresponding sample labels, is the prototype loss function between the local prototype and the global prototype of the i-th remaining client, is the local prototype of the i-th remaining client, is the global prototype, is the set weight coefficient, For the Class global prototype, Indicates that the The client collection of the class's global prototype, For the Of the remaining clients Class local prototype, K is the total number of prototype categories; According to the recovery loss function, the performance of the updated remaining client local model is optimized in combination with the gradient descent method. The expression is: ； in: For the The remaining clients perform performance optimization on the local model. is the gradient of the recovery loss function of the updated local model of the i-th remaining client on the local dataset.

7. The prototype-based federated forgetting learning method according to claim 6, characterized in that: The gradient of the recovery loss function of the updated local model of the i-th remaining client on the local dataset is specifically expressed as: 。 8. A prototype-based federated forgetting learning device, characterized in that: include: An acquisition module is used to obtain the local prototype and forget request calculated as the target prototype when the target client is ready to exit the initial federated learning training phase; Interrupt module: used to interrupt the aggregation process of the federated learning training phase and enter the forgetting phase based on the forgetting request; A processing module is used to, during the forgetting phase, enable the remaining clients to adopt a model forgetting strategy based on the target prototype to eliminate the influence of the target client data on the local models of the remaining clients, and introduce a global prototype to perform model recovery training on the local prototype after the influence is eliminated; The recovery module is used to obtain the local models of the remaining clients after the recovery training and aggregate them to obtain the reconstructed global model, and restore the initial federated learning training phase based on the reconstructed global model.

9. A prototype-based federated forgetting learning system, characterized in that include: Memory, used to store computer programs / instructions; A processor, configured to execute the computer program / instructions to implement the steps of the prototype-based federated forgetting learning method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program / instruction stored thereon, characterized in that: When the computer program / instruction is executed by a processor, the steps of the prototype-based federated forgetting learning method described in any one of claims 1-7 are implemented.

Images