A fairness-aware incentive method for federated classification learning

Abstract

The application provides a fair perception incentive method for federated classification learning, comprising the following steps: step 1, a server initializes a global prototype; step 2, a client initializes local model parameters, trains the local model parameters by combining the global prototype and local private data, and updates the local model parameters and the local prototype; step 3, the client uploads the updated local model parameters and the local prototype to the server; step 4, quality detection is performed on the received local model parameters to generate an edge incentive; step 5, the edge incentive and the updated local prototype are used for aggregation to obtain a consensus prototype; step 6, prototype-level aggregation is performed to obtain a new global prototype; and step 7, steps 2-6 are repeated until a maximum number of cycles is reached, and the final global prototype and the personalized model of each client are output. The application can better alleviate the influence of data heterogeneity and class imbalance on model accuracy, ensure cooperation fairness and encourage high-quality clients to participate in federated learning.

Description

Technical Field

[0001] This invention relates to a fair perception incentive method for federated classification learning, belonging to the field of image processing technology. Background Technology

[0002] Image classification is a crucial task in computer vision, widely applied in fields such as autonomous driving, facial recognition, and smart healthcare. These technologies rely on large-scale, high-quality labeled data to build accurate models. However, due to limitations imposed by privacy, security, and ethical concerns, data is often distributed in isolated silos, severely hindering the development and application of image classification algorithms. Federated learning has successfully addressed the challenges of large-scale data analysis and privacy protection, breaking down the barriers of data silos.

[0003] Despite the numerous advantages of federated learning, data heterogeneity and class imbalance are unavoidable in practical applications. For example, differences in imaging equipment and protocols used by different medical institutions can easily lead to data heterogeneity; simultaneously, significant variations in the number of samples for different diseases result in class imbalance. Under such data distribution, the involvement of certain clients may degrade model performance.

[0004] Depending on the optimization objective, existing solutions mainly fall into two categories: one is learning a more generalizable global model, such as by adding regularization terms during local training; the other is personalized federated learning algorithms, which learn a personalized model for each client to adapt to the local data distribution. However, these methods typically assume that each client participates in training voluntarily, ignoring the client's self-interest—that is, clients may not be willing to contribute resources unconditionally. If the model performance is poor and there is a lack of effective incentive mechanisms, rational clients may choose to withdraw from federated learning, leading to a significant drop in overall training performance. How to incentivize high-quality client participation and conduct fair contribution evaluation is crucial for the successful implementation of federated learning.

[0005] While Shapley values ​​exist and can be used to fairly evaluate client contributions in federated learning, existing Shapley-based schemes have the following limitations. First, the complexity of calculating Shapley values ​​increases exponentially with the number of clients. Second, the computation time for Shapley values ​​is heavily dependent on the size of the validation set. Finally, Shapley values ​​are calculated at the client level, failing to account for each client's contribution to each category.

[0006] Against this backdrop, this invention proposes a fairness-perceived incentive method for federated classification learning to mitigate the impact of data heterogeneity and class imbalance, and to attract more high-quality customers to participate. Currently, research on federated classification learning methods in practical applications is limited, and research with engineering applicability is even more valuable. Summary of the Invention

[0007] This invention addresses the technical problems of clients being unwilling to unconditionally contribute resources and low overall training performance in existing federated learning technologies, and proposes a fair perception incentive method for federated classification learning.

[0008] The technical solution adopted by this invention to solve the above problems is as follows: This invention proposes a fair perception incentive method for federated classification learning, comprising:

[0009] Step 1: The server initializes the global prototype and sends the initialized global prototype to the client;

[0010] Step 2: The client initializes the local model parameters, uses the received global prototype and local private data for training, and updates the local model parameters and local prototype.

[0011] Step 3: The client uploads the updated local model parameters and local prototype to the server;

[0012] Step 4: The server performs quality checks on the received local model parameters and generates edge stimuli driven by the server.

[0013] Step 5: The server aggregates the edge incentives driven by the server and the updated local prototype to obtain the consensus prototype;

[0014] Step 6: Perform prototype-level aggregation based on the formula prototype to obtain a new global prototype;

[0015] Step 7: Repeat steps 2-6 until the maximum number of iterations is reached, and output the final global prototype and the personalized models for each client W = {w1, L, w N}

[0016] Optionally, step 4, which generates the server-driven edge excitation, includes:

[0017] Step 4.1: The client executes the received local model parameters and local prototype, and calculates the loss value of the local prototype of the i-th client. and the loss value of the global prototype

[0018] Step 4.2: If the performance of the local prototype is lower than that of the global prototype, the server will provide negative incentives to the corresponding client and remove it;

[0019] Step 4.3: If the local prototype outperforms the global prototype, the server provides positive incentives to the corresponding client, and the edge incentive range is normalized during each round of local model parameter quality detection.

[0020] The formula for calculating edge excitation is:

[0021]

[0022] In formula (1), To calculate the loss value of the local prototype for the i-th client, Let be the loss value of the global prototype, and ζ(·) be the activation function. For edge excitation, if Apply negative incentives, if Provide positive incentives.

[0023] Optionally, the expression for the consensus prototype in step 5 is:

[0024]

[0025] In formula (2), For edge excitation, This is a local prototype.

[0026] Optionally, step 6, obtaining the consensus prototype, includes:

[0027] Step 6.1: The server estimates the Shapley value of the participants based on the consensus prototype, and evaluates the participants' contributions at the prototype level by using the estimated Shapley value;

[0028] Step 6.2: The server aggregates the participants' contributions at the prototype level to obtain a new global model.

[0029] Optionally, step 6.1, which involves estimating the Shapley value based on the consensus prototype, includes:

[0030] Step 6.1.1: Calculate the aggregate prototype of class k that does not contain client i.

[0031] Step 6.1.2: Normalize the contributions of different types of prototypes in each client at the prototype level;

[0032] Step 6.1.3: Obtain the k-th class prototype of client i. For the aggregate prototype of class k that does not contain client i The contribution is used as an estimate of the Shapley value;

[0033] The aggregate prototype of class k that does not contain client i The calculation formula is:

[0034]

[0035] The normalized expression is:

[0036]

[0037] In formula (4), S t Let be the set of clients that receive positive incentives in round t;

[0038] The expression for Shapley value estimation is:

[0039]

[0040] In formula (5), This is an estimate of the Shapley value.

[0041] Optionally, the expression for prototype-level aggregation in step 6.2 is:

[0042]

[0043] In formula (6), P G,K This is the new global prototype.

[0044] The beneficial effects of this invention are: compared with the prior art, this invention can better mitigate the impact of data heterogeneity and class imbalance on model accuracy, ensure fair collaboration and incentivize high-quality customers to participate in federated learning, thereby obtaining better global prototype and model performance. Attached Figure Description

[0045] Figure 1 A flowchart illustrating a fair perception incentive method for federated classification learning provided by this invention;

[0046] Figure 2 This is a schematic diagram illustrating the local client training and server-side incentive calculation process provided by the present invention.

[0047] Figure 3 A comparison of the BACC (%) curves of a fair perception incentive method for federated classification learning provided by this invention and seven existing benchmark methods. Example

[0048] Combination Figure 1-3 This embodiment will be described as follows: Figure 1 As shown in this embodiment, the process of a fairness-perceived incentive method for federated classification learning includes:

[0049] S1: Server-side initialization of global prototype Client initializes local model parameters The local models of each client can be heterogeneous;

[0050] S2: The server distributes the global prototype to each client.

[0051] S3: As Figure 2 As shown, the client performs local training;

[0052] S301: For the t-th communication, the client utilizes local private data D i and global prototype For local training, the loss function is as follows:

[0053]

[0054] In formula (1), L i Let y represent the local training loss function for the i-th client, K represent the total number of classes, and y represent the loss function for the i-th client. j,k Let represent the true label of the j-th sample in the training set corresponding to the k-th class, and log(·) denote the logarithmic operation with base e. This represents the i-th client model parameter. Let x represent the j-th sample. j The input to the i-th client-side local model represents the probability of belonging to the k-th class. λ represents a hyperparameter, which is set to 1 in this experiment. Let represent the local prototype of the i-th client with category k in round t.

[0055] S302: Local Prototype It is derived from the training image x j The feature vector f obtained by inputting into the network feature extractor i (x j ), where d is the dimension of the feature vector, obtained by averaging the feature vectors of the k-th class;

[0056] The formula for calculating the local prototype is:

[0057]

[0058] In formula (2), D i,k Let represent the sample of the i-th client that contains category k.

[0059] S4: As Figure 2 As shown, the client will use the trained local model parameters. and local prototype The data is uploaded to the server, which uses the validation set to calculate the loss value of the local prototype for the i-th client. and the loss value of the global prototype

[0060] S5: The server determines whether the local prototype is superior to the global prototype.

[0061] Based on server-side judgment: If the above conditions are met, it indicates that the local prototype is superior to the global prototype, and the server will provide positive incentives to the client to encourage its active participation in federated learning; otherwise, zero or negative incentives will be provided to filter out these clients. The formula for calculating server-driven edge incentives is:

[0062]

[0063] In formula (3), This represents the edge excitation driven by the server, where ζ(·) is the activation function, which can be selected according to different classification tasks. Since... The range gradually decreases as training progresses, and normalization is performed during each round of computation. This design effectively filters out low-quality customers and customers with significantly different data distributions, thereby reducing the overall performance degradation they cause.

[0064] S6: Aggregate server-driven edge incentives and local prototypes to obtain a high-quality consensus prototype:

[0065]

[0066] In formula (4), Let S represent the consensus prototype obtained in round t. t Let represent the set of clients that receive positive incentives in round t.

[0067] S7: The server performs prototype-based Shapley value estimation on the formula prototype.

[0068] The server evaluates the contributions of different client prototype categories to ensure collaborative fairness. This not only accelerates Shapley calculations but also eliminates reliance on the size of the validation set and allows for hierarchical estimation. The server uses the contribution of the i-th client's k-th class prototype to the aggregated prototypes of k classes that do not include that client in the co-prototype as its utility function. The Shapley utility function for the i-th client's k-th class prototype is shown below. as follows:

[0069]

[0070] In formula (5), This represents the aggregate prototype of class k, excluding client i.

[0071] Considering the varying contributions of different types of prototypes across clients, normalization was further performed at the prototype level to ensure that the sum of clients containing the k-th type of prototype is 1.

[0072] S8: Server-side aggregation at the prototype level:

[0073]

[0074] Unlike client-level aggregation, prototype-level aggregation can focus on various categories, making it a more granular aggregation method. This not only mitigates the impact of category imbalance but also ensures fairness in collaboration.

[0075] S9: Repeat S2-S8 until the preset training stopping condition is met, obtaining the global prototype and the personalized models W = {w1,L,w...} for each client. N}

[0076] In summary, to verify the classification accuracy of this invention, the following experiments were conducted:

[0077] First, we introduce the actual medical image dataset used in the experiment: the dermoscopy dataset Fed-ISIC2019. This dataset contains 23,247 images, divided into six clients, covering eight types of skin lesions. The image size is 200×200×3. We used FedAvg, FedProx, FedBN, FedProto, INCFL, ShapleyFL, and FedCE as benchmark methods. We used BACC (balanced accuracy) as the classification metric. As shown in Table 1, the federated image classification method proposed in this invention achieves high classification accuracy, improving upon FedAvg, FedProx, FedBN, FedProto, INCFL, ShapleyFL, and FedCE by 16.59%, 19.44%, 10.74%, 4.23%, 15.49%, 13.61%, and 7.61%, respectively.

[0078] Table 1

[0079] 50.77 47.92 56.62 63.13 51.87 53.75 59.75 67.36

[0080] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent substitutions, and improvements made to the above embodiments without departing from the scope of the present invention, based on the technical essence of the present invention and within the spirit and principles of the present invention, shall still fall within the protection scope of the present invention.

Claims

1. A fairness-perceived incentive method for federated classification learning, characterized in that, The steps of the fairness-perceived incentive method for federated classification learning include: Step 1: The server initializes the global prototype and sends the initialized global prototype to the client; Step 2: The client initializes the local model parameters, uses the received global prototype and local private data for training, and updates the local model parameters and local prototype. Step 3: The client uploads the updated local model parameters and local prototype to the server; Step 4: The server performs quality checks on the received local model parameters and generates edge stimuli driven by the server. Step 4, which generates server-driven edge stimuli, includes the following steps: Step 4.1: The client executes the received local model parameters and local prototype to calculate the first... i Loss value of the local prototype on each client and the loss value of the global prototype ; Step 4.2: If the performance of the local prototype is lower than that of the global prototype, the server will provide negative incentives to the corresponding client and remove it; Step 4.3: If the local prototype outperforms the global prototype, the server provides positive incentives to the corresponding client, and the edge incentive range is normalized during each round of local model parameter quality detection. The formula for calculating edge excitation is: （1） In formula (1), To calculate the first i The loss value of the local prototype for each client. This represents the loss value for the global prototype. For activation function, For edge excitation, if If negative incentives are applied, Provide positive incentives; Step 5: The server aggregates the server-driven edge incentives and the updated local prototype to obtain the consensus prototype; Step 6: Perform prototype-level aggregation based on the formula prototype to obtain a new global prototype; Step 7: Repeat steps 2-6 until the maximum number of loops is reached, and output the final global prototype and the personalized models for each client. .

2. The fairness-perceived incentive method for federated classification learning according to claim 1, characterized in that, The expression for the consensus prototype in step 5 is: （2） In formula (2), For edge excitation, This is a local prototype.

3. The fairness-perceived incentive method for federated classification learning according to claim 1, characterized in that, Step 6, obtaining the consensus prototype, includes the following steps: Step 6.1: The server estimates the Shapley value of the participants based on the consensus prototype, and evaluates the participants' contributions at the prototype level by using the estimated Shapley value; Step 6.2: The server aggregates the participants' contributions at the prototype level to obtain a new global model.

4. The fairness-perceived incentive method for federated classification learning according to claim 3, characterized in that, Step 6.1, which involves estimating the Shapley value based on the consensus prototype, includes the following steps: Step 6.1.1: Calculate the number of cases that do not include clients. i of k Aggregate prototype of a class ; Step 6.1.2: Normalize the contributions of different types of prototypes in each client at the prototype level; Step 6.1.3: Obtain the k-th class prototype of client i. For those that do not include clients i of k Aggregate prototype of a class The contribution is used as an estimate of the Shapley value; Not including client i of k Aggregate prototype of a class The calculation formula is: （3） The normalized expression is: （4） In formula 4, Let be the set of clients that receive positive incentives in round t; The expression for Shapley value estimation is: （5) In formula (5), This is an estimate of the Shapley value.

5. A fairness-perceived incentive method for federated classification learning according to claim 3, characterized in that, The expression for prototype-level aggregation in step 6.2 is: （6） In formula (6), This is the new global prototype.

Images