Federal learning security excitation method, device and system based on multi-dimensional evaluation

Abstract

The invention discloses a federated learning security excitation method, device and system based on multi-dimensional evaluation. The method comprises the following steps: issuing a federated learning task and distributing an initialized global model to each client; receiving the data volume, the information entropy and the masked model parameters uploaded by each client; carrying out mask removal processing on the model parameters of the masked clients to obtain the model parameters; calculating to obtain a data volume contribution degree, an information entropy contribution degree and a marginal contribution degree of each client according to the data volume, the information entropy and the model parameters of each client; determining a dynamic weight of the client; performing weighted aggregation on the model parameters according to the dynamic weight of each client to generate new global model parameters; and issuing rewards to the clients according to the dynamic weights of the clients. By using the scheme of the invention, the security in the uploading process of the client model parameters can be ensured, and each client can obtain fair and reasonable excitation.

Description

Technical Field

[0001] This application relates to the field of federated learning, and in particular to a method, apparatus, and system for providing secure incentives for federated learning based on multi-dimensional assessment. Background Technology

[0002] With the rapid development of artificial intelligence technology, data is playing an increasingly prominent role as a core driving force. However, data privacy protection and the problem of data silos have become bottlenecks restricting cross-institutional collaboration. Federated Learning (FL), through its distributed machine learning training paradigm of "data remains stationary while the model moves," provides a feasible path to break down data silos. Its core objective is to allow multiple parties to collaborate in building a global model without sharing the original data, and it has already shown application potential in fields such as financial risk control, smart healthcare, and the Industrial Internet of Things. However, the large-scale implementation of federated learning still faces two key challenges: insufficient fairness in incentive mechanism design and privacy and security threats.

[0003] Regarding incentive mechanisms, existing solutions often use a single indicator (such as data volume or local training rounds) to quantify participant contributions, which is difficult to adapt to the heterogeneity and quality differences of data in real-world scenarios. For example, while the traditional Shapley value method can reflect marginal contributions, it does not consider the informational value and dynamic changes of data, leading to an imbalance in resource allocation and reducing the incentive for high-quality participants. Research shows that incentive mechanisms relying solely on data volume may lead to a "free-rider" phenomenon, where low-quality participants receive high rewards due to large data volumes, while high-quality contributors with small amounts of data are underestimated. This problem is particularly prominent in privacy-sensitive fields such as healthcare and finance, directly impacting the sustainable development of collaborative ecosystems.

[0004] From a security perspective, the distributed architecture of federated learning exposes it to multiple attack risks: 1) Privacy leakage risk: Malicious servers or participants can steal original data features through reverse inference of model parameters (such as member inference attacks); 2) Model attack threat: Byzantine clients may upload tampered gradients or parameters (such as backdoor attacks or model poisoning), damaging global model performance. Although existing research has proposed protective measures such as differential privacy and homomorphic encryption, these often come at the cost of sacrificing model accuracy or increasing computational overhead. For example, homomorphic encryption doubles the communication load, while differential privacy technology reduces model convergence speed. How to maintain model availability while efficiently protecting privacy remains a critical technical challenge that needs to be overcome. Summary of the Invention

[0005] This application provides a method, apparatus, and system for secure incentives in federated learning based on multi-dimensional evaluation, in order to address the fairness deficiencies of incentive mechanisms and the security issues of privacy data in traditional federated learning systems.

[0006] Therefore, this application provides the following technical solution: This application provides a federated learning security incentive method based on multi-dimensional evaluation, the method comprising: Publish the federated learning task and distribute the initialized global model to each client; The system receives the amount of data, information entropy, and masked model parameters uploaded by each client; the masked model parameters are obtained by adding a mask to the model parameters trained by the client on local data for the initial global model. The masked model parameters for each client are demasked to obtain the model parameters; The data volume contribution, information entropy contribution, and marginal contribution of each client are calculated based on the data volume, information entropy, and model parameters of each client, respectively. The dynamic weight of each client is determined based on its data volume contribution, marginal contribution, and information entropy contribution. The masked model parameters are weighted and aggregated based on the dynamic weights of each client to generate new global model parameters; Rewards are distributed to the clients based on their dynamic weights. The data volume contribution rate is used to measure the relative proportion of client data volume, and its calculation formula is as follows: ; in, N This indicates the number of clients participating in federated learning; D i Indicates the amount of data in client i; The marginal contribution refers to the client's contribution to the improvement of the global model's performance. The marginal contribution of the client to the global model is calculated based on the Shapley value, i.e.: (3) in, N This represents the number of clients participating in federated learning; S represents a subset of clients, i.e., a subset selected from all clients, numbered from 1 to N. U(S) The utility of the model when a subset S is used for training, including the accuracy of the model parameters; The information entropy contribution indicates that the higher the data quality, the greater the entropy and the higher the weight ratio.

[0007] The formula for calculating the contribution of information entropy is as follows: ; in, N This indicates the number of clients participating in federated learning; H ( D i) indicates the client i Uploaded information entropy.

[0008] Optionally, the federated learning task includes: task objective, parameter requirements, and stopping conditions.

[0009] Optionally, different clients may have their own independent local data, and the local data of different clients may have different distributions and characteristics.

[0010] Optionally, determining the dynamic weight of a client based on its data volume contribution, marginal contribution, and information entropy contribution includes: performing a weighted calculation on the client's data volume contribution, marginal contribution, and information entropy contribution to obtain the client's dynamic weight.

[0011] Optionally, the method further includes: The client receives the federated learning task and initializes the global model; The initialized global model is trained on local data according to the federated learning task to obtain model parameters; Calculate the information entropy of the local data and add a mask to the model parameters to obtain the masked model parameters; Upload the information entropy and the masked model parameters.

[0012] Optionally, calculating the information entropy of the local data includes:

[0013] in K The total number of categories, where each category refers to a different label or grouping in the dataset used for model training; n k For category k The number of samples; | D i | represents the total amount of data.

[0014] Optionally, adding a mask to the model parameters to obtain masked model parameters includes: Generate a random mask; The random mask is added to the model parameters to obtain the masked model parameters.

[0015] This application also provides a federated learning security incentive device based on multi-dimensional evaluation, the device comprising: The publishing module is used to publish federated learning tasks and distribute the initialized global model to each client. The receiving module is used to receive the amount of data, information entropy, and masked model parameters uploaded by each client; the masked model parameters are obtained by adding a mask to the model parameters obtained by the client training the initial global model on local data. The data processing module is used to demask the masked model parameters of each client to obtain the model parameters; The contribution calculation module is used to calculate the data volume contribution, information entropy contribution, and marginal contribution of each client based on the data volume, information entropy, and model parameters of each client, respectively. The weight calculation module is used to determine the dynamic weight of each client based on the data volume contribution, the marginal contribution, and the information entropy contribution of each client. The aggregation module is used to perform weighted aggregation of the masked model parameters based on the dynamic weights of each client, and generate new global model parameters. The reward module is used to distribute rewards to the clients based on their dynamic weights.

[0016] This application embodiment also provides a federated learning security incentive system based on multi-dimensional evaluation. The system includes: a server and at least two clients. The server includes the federated learning security incentive device based on multi-dimensional evaluation. The server is used to publish federated learning tasks and initialize a global model to each client, determine the dynamic weight of each client based on the amount of data uploaded by the client, information entropy, and masked model parameters, perform weighted aggregation of the model parameters of each client based on the dynamic weight, generate new global model parameters, and distribute rewards to the clients based on the dynamic weight. The client is used to calculate the information entropy of local data, train the initial global model published by the server based on the local data to obtain model parameters, add a mask to the model parameters, and upload the masked model parameters, data volume, and information entropy to the server.

[0017] Optionally, the client includes: The receiving unit is used to receive the federated learning task and initialize the global model; The training unit is used to train the initialized global model on local data according to the federated learning task to obtain model parameters; A fuzzing unit is used to calculate the information entropy of the local data and add a mask to the model parameters to obtain masked model parameters. The upload unit is used to upload the information entropy and the masked model parameters.

[0018] This application also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, performs the steps of the federated learning security incentive method based on multi-dimensional evaluation.

[0019] The federated learning security incentive method, apparatus, and system based on multi-dimensional evaluation provided in this application protect the privacy of client model parameters during the upload process and prevent inference attacks and Byzantine attacks by using an eliminateable random masking perturbation technique. By evaluating the client's contribution from multiple dimensions and dynamically adjusting the weights, the incentive mechanism can be made fairer and more reasonable, effectively incentivizing clients to actively contribute high-quality data and improving the performance and data availability of the global model. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly described below. Obviously, the drawings described below are merely some embodiments of the present invention, and those skilled in the art can obtain other drawings based on these drawings without creative effort.

[0021] Figure 1 This is a flowchart of a federated learning security incentive method based on multi-dimensional evaluation provided in an embodiment of this application; Figure 2 This is a schematic diagram of a structure of a federated learning security incentive device based on multi-dimensional evaluation provided in an embodiment of this application; Figure 3 This is a schematic diagram of a federated learning security incentive system based on multi-dimensional evaluation provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of a client structure in an embodiment of the present invention. Detailed Implementation

[0022] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0023] To address the limitations of traditional federated learning, where participant contribution evaluation is confined to a single dimension (such as data volume), making it difficult to adapt to quality differences in heterogeneous data environments, and the impact on system reliability caused by privacy leaks and malicious attacks during model parameter transmission, this application provides a federated learning security incentive method, apparatus, and system based on multi-dimensional evaluation. By dynamically evaluating data volume, marginal contribution, and information value from multiple dimensions, it solves the fairness defects of traditional incentive mechanisms; and it designs an eliminable random mask perturbation technique to protect the accuracy of model aggregation while resisting inference attacks and Byzantine attacks.

[0024] like Figure 1 The diagram shown is a flowchart of a federated learning security incentive method based on multi-dimensional evaluation provided in this application, which includes the following steps: In step 101, the federated learning task is published and the initialized global model is distributed to each client.

[0025] The specific structure and initial model parameters of the global model can be determined by the server according to the actual application needs, and this embodiment of the invention does not limit this.

[0026] The federated learning task may include information such as task objectives, parameter requirements, and stopping conditions.

[0027] In this embodiment of the invention, each client holds a local dataset, meaning that each client has its own independent data, which may have different distributions and characteristics.

[0028] Accordingly, after receiving the federated learning task and initializing the global model from the server, each client performs local model training. It trains the model on its own local data using a suitable machine learning algorithm to obtain local model parameters. This embodiment of the invention does not limit the algorithm used for model training; it can be selected according to the actual application needs.

[0029] Furthermore, each client also needs to calculate the information entropy of its local data and provide it to the server. Information entropy is a core indicator used to quantify information uncertainty; it reflects the average degree of uncertainty of the information source.

[0030] Furthermore, to ensure confidentiality and resistance to attacks during model data upload and to guarantee data security, each client can add a random mask to the model parameters before uploading. It should be noted that these masks should be removable so that the server can recover the original model parameters. The method of perturbation through masks can generate removable mask vectors and significantly reduce time overhead; therefore, this invention employs masking technology for privacy protection. For example, client iUse local data D i Training the model yields model parameters And calculate information entropy Information entropy The calculation formula is as follows: (1) in, K The total number of categories refers to the different labels or groups in the dataset used for model training, which represent different features or attributes of the samples. n k For category k The number of samples; | D i | represents the total amount of data.

[0031] In this embodiment of the invention, a higher information entropy indicates a more balanced data distribution and a greater amount of information.

[0032] In this embodiment of the invention, the client can add masks to the model parameters in various ways. For example, in a non-limiting embodiment, the client... i Random mask vectors can be generated collaboratively with a server using the Diffie-Hellman protocol. Add the random mask to the model parameters In the process, the masked model parameters are generated, namely: .

[0033] Accordingly, when the server performs weighted aggregation on the masked model parameters of each client, it is necessary to first remove the mask from the model parameters before performing weighted aggregation.

[0034] Of course, other masking methods can also be used, as long as the server can correctly remove the mask.

[0035] Client-side data upload volume | D i Information entropy and masked model parameters Give it to the server.

[0036] Accordingly, in step 102, the amount of data uploaded by each client, the information entropy, and the masked model parameters are received.

[0037] In step 103, the masked model parameters of each client are demasked to obtain the model parameters.

[0038] As mentioned earlier, there are multiple ways to mask model parameters, and correspondingly, when performing support masking, the corresponding demasking method can be used.

[0039] After receiving the model parameters, the server obtains the model parameters through a demasking operation. .

[0040] In step 104, the data volume contribution, information entropy contribution, and marginal contribution of each client are calculated based on the data volume, information entropy, and model parameters of each client.

[0041] In federated learning, three metrics—data volume, marginal contribution, and information entropy—work together to comprehensively evaluate client contributions: Data volume reflects the relative scale of client data resources, ensuring that participants with large data volumes receive appropriate weights; marginal contribution quantifies the actual improvement a client makes to model performance through Shapley value, reflecting its strategic value; and information entropy measures data quality—higher entropy values ​​indicate greater data diversity and richer information content, making it more crucial for improving model generalization ability. The combination of these three metrics avoids biases caused by solely relying on data volume and identifies high-quality data and efficient participants, thus achieving a fair and dynamic weight allocation.

[0042] Therefore, in this embodiment of the invention, the data volume contribution, marginal contribution, and information entropy contribution of the client can be determined respectively; then, the dynamic weight of the client is calculated by combining these different dimensions of contribution.

[0043] The calculation methods for the contribution of each dimension are explained below.

[0044] (1) Data volume contribution: This measures the relative proportion of client data volume, and its calculation formula is as follows: (2) in, N This indicates the number of clients participating in federated learning.

[0045] (2) Marginal contribution: refers to the contribution of the client to the improvement of the global model performance.

[0046] Specifically, the marginal contribution of the client to the global model can be calculated based on the Shapley value, i.e.: (3) in, N This represents the number of clients participating in federated learning; S represents a subset of clients, i.e., a subset selected from all clients (numbered 1 to N); U(S) The utility of the model when a subset S is used for training (such as the accuracy of model parameters).

[0047] The above formula is used to calculate the first... i The marginal contribution of each client to the global model. Specifically, it considers all clients that do not include the first client.i A subset of clients S And calculated the number of times the first... i The increase in model utility when each client joins these subsets is calculated. This increase is assigned a weight, and the weighted sum of all possible increases yields the result for the nth client. i Marginal contribution of each client.

[0048] (3) Information entropy contribution: The higher the data quality (the greater the entropy), the higher the weight ratio.

[0049] The formula for calculating the contribution of information entropy is as follows: (4) in, N This indicates the number of clients participating in federated learning; H ( D i ) indicates the client i Uploaded information entropy.

[0050] In step 105, the dynamic weight of each client is determined based on its data volume contribution, marginal contribution, and information entropy contribution.

[0051] Based on the above three dimensions, a weighted calculation is performed to obtain the dynamic weight of each client. w i The calculation formula is as follows: (5) in, α , β , γ For the weighting coefficients, satisfying α + β + γ =1.

[0052] It should be noted that the coefficient α , β , γ It can be dynamically optimized using existing entropy weighting methods or reinforcement learning. Specifically: The entropy weight method eliminates the influence of dimensions by standardizing data, calculates the information entropy of each indicator to measure the degree of dispersion, and determines the objective weight based on the difference in entropy value to reflect the difference in the importance of the indicators.

[0053] Reinforcement learning simulates the interaction between the environment and the agent. The agent adjusts its policy based on the rewards received from the environment to maximize cumulative rewards. When optimizing coefficients, each coefficient can be treated as an action, and these coefficients can be dynamically adjusted using reinforcement learning algorithms (such as Q-learning or policy gradient methods) to find the optimal combination.

[0054] In step 106, the masked model parameters are weighted and aggregated according to the dynamic weights of each client to generate new global model parameters.

[0055] The masked model parameters are weighted and aggregated to obtain the aggregated model of all participating clients, which is the new global model parameter. : .

[0056] By weighting and aggregating the model parameters of each client based on dynamic weights, the parameters of high-weight clients can have a greater impact on the global model, while the impact of low-weight or abnormal clients can be suppressed.

[0057] By aggregating the model parameters from each client, a more representative and generalizable global model can be generated.

[0058] In step 107, rewards are distributed to the clients based on their dynamic weights.

[0059] Specifically, rewards can be sent proportionally based on the dynamic weight of each client. These rewards can be currencies, tokens, or resource quotas, etc., and this embodiment of the invention does not limit the specific rewards.

[0060] Furthermore, the server can add new global model parameters. Distributed to all clients; clients receive global model parameters. Then use these as initial parameters for the next round of local training; repeat the above steps until the model converges.

[0061] The federated learning security incentive method based on multi-dimensional evaluation provided in this invention ensures the security of client model parameters during the upload process through an eliminable random mask perturbation technique. By evaluating the client's contribution from multiple dimensions and dynamically adjusting the weights, the incentive mechanism can be made fairer and more reasonable, thereby increasing the client's enthusiasm and improving the performance and data availability of the global model.

[0062] Accordingly, embodiments of the present invention also provide a federated learning security incentive device based on multi-dimensional evaluation, such as... Figure 2 The diagram shown is a structural schematic of the device.

[0063] The federated learning security incentive device 200, based on multi-dimensional assessment, includes the following modules: Module 201 is used to publish federated learning tasks and distribute the initialized global model to each client; The receiving module 202 is used to receive the amount of data, information entropy, and masked model parameters uploaded by each client; the masked model parameters are obtained by adding a mask to the model parameters obtained by the client training the initial global model on local data; Data processing module 203 is used to demask the masked model parameters of each client to obtain the model parameters; The contribution calculation module 204 is used to calculate the data volume contribution, information entropy contribution, and marginal contribution of each client based on the data volume, information entropy, and model parameters of each client, respectively. The weight calculation module 205 is used to determine the dynamic weight of each client based on the data volume contribution, the marginal contribution, and the information entropy contribution of each client. The aggregation module 206 is used to perform weighted aggregation of the masked model parameters according to the dynamic weights of each client, and generate new global model parameters. The reward module 207 is used to distribute rewards to the clients based on the dynamic weight of each client.

[0064] The specific implementation methods of the above modules can be referred to the description in the previous embodiment of the federated learning security incentive method based on multi-dimensional evaluation of the present invention, and will not be repeated here.

[0065] Accordingly, embodiments of the present invention also provide a federated learning security incentive system based on multi-dimensional evaluation, such as... Figure 3 The diagram shown is a structural schematic of the system.

[0066] This federated learning security incentive system, based on multi-dimensional evaluation, comprises: 30 servers and at least two clients 40, wherein: Server 30 includes the aforementioned federated learning security incentive device 200 based on multi-dimensional evaluation, which is used to issue federated learning tasks and initialize global models to each client 40, determine the dynamic weight of each client 40 based on the amount of data uploaded by the client 40, information entropy and masked model parameters, perform weighted aggregation of the model parameters of each client 40 based on the dynamic weight, generate new global model parameters, and issue rewards to the client 40 based on the dynamic weight. Client 40 is used to calculate the information entropy of local data and train the initial global model published by server 30 based on the local data to obtain model parameters. The model parameters are then masked, and the masked model parameters, data volume, and information entropy are uploaded to server 30.

[0067] like Figure 4 The diagram shown is a structural schematic of a client in an embodiment of the present invention.

[0068] In this embodiment, the client 40 includes the following units: The receiving unit 401 is used to receive federated learning tasks and initialize the global model; Training unit 402 is used to train the initialized global model on local data according to the federated learning task to obtain model parameters; The fuzzing unit 403 is used to calculate the information entropy of the local data and add a mask to the model parameters to obtain the masked model parameters. The upload unit 404 is used to upload the data volume, the information entropy, and the masked model parameters. The data volume refers to the amount of local data on the client 40.

[0069] It should be understood that, although Figure 1 The steps in the flowchart shown are displayed sequentially as indicated by the arrows; however, these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise explicitly stated herein, there is no strict order in which these steps are executed, and they can be performed in other orders. Furthermore, Figure 1 At least some of the steps in the process may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but may be executed at different times. The execution order of these steps or stages is not necessarily sequential, but may be executed in turn or alternately with other steps or at least some of the steps or stages in other steps.

[0070] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the methods described above. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, or optical storage, etc. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.

[0071] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0072] The embodiments of this application have been described in detail above. Specific implementation methods have been used to illustrate this application. The descriptions of the embodiments above are only for the purpose of helping to understand the methods and systems of this application, and are merely a part of the embodiments of this application, not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this application, and the content of this specification should not be construed as a limitation of this application. Therefore, any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A federated learning security incentive method based on multi-dimensional evaluation, characterized in that, The method includes: Publish the federated learning task and distribute the initialized global model to each client; The system receives the amount of data, information entropy, and masked model parameters uploaded by each client; the masked model parameters are obtained by adding a mask to the model parameters trained by the client on local data for the initial global model. The masked model parameters for each client are demasked to obtain the model parameters; The data volume contribution, information entropy contribution, and marginal contribution of each client are calculated based on the data volume, information entropy, and model parameters of each client, respectively. The dynamic weight of each client is determined based on its data volume contribution, marginal contribution, and information entropy contribution. The masked model parameters are weighted and aggregated based on the dynamic weights of each client to generate new global model parameters; Rewards are distributed to the clients based on their dynamic weights. The data volume contribution rate is used to measure the relative proportion of client data volume, and its calculation formula is as follows: ； in, N This indicates the number of clients participating in federated learning; D i Indicates the amount of data in client i; The marginal contribution refers to the client's contribution to the improvement of the global model's performance. The marginal contribution of the client to the global model is calculated based on the Shapley value, i.e.: in, N This represents the number of clients participating in federated learning; S represents a subset of clients, i.e., a subset selected from all clients, numbered from 1 to N. U(S) The utility of the model when a subset S is used for training, including the accuracy of the model parameters; The information entropy contribution indicates that the higher the data quality, the greater the entropy and the higher the weight ratio. The formula for calculating the contribution of information entropy is as follows: ； in, N This indicates the number of clients participating in federated learning; H ( D i ) indicates the client i Uploaded information entropy.

2. The federated learning security incentive method based on multi-dimensional evaluation according to claim 1, characterized in that, The federated learning task includes: task objectives, parameter requirements, and stopping conditions.

3. The federated learning security incentive method based on multi-dimensional evaluation according to claim 2, characterized in that, Different clients have their own independent local data, and the local data of different clients have different distributions and characteristics.

4. The federated learning security incentive method based on multi-dimensional evaluation according to claim 1, characterized in that, The determination of the dynamic weight of a client based on its data volume contribution, marginal contribution, and information entropy contribution includes: The dynamic weight of the client is obtained by weighting its data volume contribution, marginal contribution, and information entropy contribution.

5. The federated learning security incentive method based on multi-dimensional evaluation according to any one of claims 1 to 4, characterized in that, The method further includes: The client receives the federated learning task and initializes the global model; The initialized global model is trained on local data according to the federated learning task to obtain model parameters; Calculate the information entropy of the local data and add a mask to the model parameters to obtain the masked model parameters; Upload the information entropy and the masked model parameters.

6. The federated learning security incentive method based on multi-dimensional evaluation according to claim 5, characterized in that, The calculation of the information entropy of the local data includes: in K The total number of categories, where each category refers to a different label or grouping in the dataset used for model training; n k For category k The number of samples; | D i | represents the total amount of data.

7. The federated learning security incentive method based on multi-dimensional evaluation according to claim 5, characterized in that, The process of adding a mask to the model parameters to obtain masked model parameters includes: Generate a random mask; The random mask is added to the model parameters to obtain the masked model parameters.

8. A federated learning security incentive device based on multi-dimensional evaluation, characterized in that, The device includes: The publishing module is used to publish federated learning tasks and distribute the initialized global model to each client. The receiving module is used to receive the amount of data, information entropy, and masked model parameters uploaded by each client; the masked model parameters are obtained by adding a mask to the model parameters obtained by the client training the initial global model on local data. The data processing module is used to demask the masked model parameters of each client to obtain the model parameters; The contribution calculation module is used to calculate the data volume contribution, information entropy contribution, and marginal contribution of each client based on the data volume, information entropy, and model parameters of each client, respectively. The weight calculation module is used to determine the dynamic weight of each client based on the data volume contribution, the marginal contribution, and the information entropy contribution of each client. The aggregation module is used to perform weighted aggregation of the masked model parameters based on the dynamic weights of each client, and generate new global model parameters. The reward module is used to distribute rewards to the clients based on their dynamic weights.

9. A federated learning security incentive system based on multi-dimensional evaluation, characterized in that, The system includes: a server and at least two clients, wherein the server includes the federated learning security incentive device based on multi-dimensional evaluation as described in claim 8; The server is used to publish federated learning tasks and initialize a global model to each client, determine the dynamic weight of each client based on the amount of data uploaded by the client, information entropy, and masked model parameters, perform weighted aggregation of the model parameters of each client based on the dynamic weight, generate new global model parameters, and distribute rewards to the clients based on the dynamic weight. The client is used to calculate the information entropy of local data, train the initial global model published by the server based on the local data to obtain model parameters, add a mask to the model parameters, and upload the masked model parameters, data volume, and information entropy to the server.

10. The federated learning security incentive system based on multi-dimensional evaluation according to claim 9, characterized in that, The client includes: The receiving unit is used to receive the federated learning task and initialize the global model; The training unit is used to train the initialized global model on local data according to the federated learning task to obtain model parameters; A fuzzing unit is used to calculate the information entropy of the local data and add a mask to the model parameters to obtain masked model parameters. The upload unit is used to upload the information entropy and the masked model parameters.

11. A computer-readable storage medium having a computer program stored thereon, characterized in that, The computer program, when run by a processor, performs the steps of the federated learning security incentive method based on multi-dimensional evaluation as described in any one of claims 1 to 7.

Images