Federal learning client data selection method based on game theory
By constructing a game theory model, the client autonomously selects a subset of data under the incentive of scarcity, which solves the problems of training efficiency and stability under heterogeneous and dynamic data distribution in federated learning, and achieves global data balance and model optimization.
Patent Information
- Application Number
- CN202511484319.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-10-17
- Publication Date
- 2026-01-09
Smart Images
Figure CN121303384A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of computer service technology and relates to a data selection method, specifically a client-side data selection method based on game theory under conditions of heterogeneous and dynamically changing data distribution. Background Technology
[0002] Federated Learning (FL) is a distributed learning paradigm that protects data privacy and is widely used in sensitive data scenarios such as finance, healthcare, and smart manufacturing. However, in practical applications, the data distribution of different clients is often heterogeneous (non-IID) and changes dynamically over time, posing significant challenges to the training efficiency, stability, and generalization ability of the global model. Existing methods mostly rely on the unified scheduling of a central server or strong assumptions about prior distribution, making it difficult to operate effectively in distributed, dynamic, and privacy-sensitive environments.
[0003] To address the aforementioned issues, if a mechanism can be designed that allows the client to autonomously choose local data with higher "global value" for training without exposing the original data, and evolve towards an equilibrium state of global data distribution under the guidance of a game mechanism, the overall performance and stability of federated learning will be significantly improved. Summary of the Invention
[0004] The purpose of this invention is to provide a game theory-based method for client data selection in federated learning, which improves the global data balance and model training stability of federated learning. This method constructs a non-cooperative game model among clients, guiding each client to autonomously contribute its data subset. Simultaneously, the server provides a global "scarcity signal" as data pricing feedback, thereby achieving a continuous balance in the global data distribution.
[0005] The objective of this invention is achieved through the following technical solution:
[0006] A game theory-based method for selecting data for federated learning clients includes the following steps:
[0007] Step S1, Game Theory Model Construction:
[0008] The client's data selection behavior is abstracted into a multi-agent non-cooperative game model. Before each round of federated training, the client independently selects a subset of data for training in that round based on local observations and feedback from the server. This game model consists of the following components:
[0009] Participants: The set of clients currently active in federated learning. .
[0010] Client status information: for each client Have local dataset The corresponding data distribution is ;
[0011] Action space: The client's policy space is all possible subsets of its local dataset, i.e., selecting a subset. This serves as the basis for this round of upload training;
[0012] Utility function: The client's benefit considers two factors: first, the contribution of the local data subset to the performance of the local model; and second, the support for the overall distribution balance of the system. This function is supported by the scarcity incentive signal provided by the central server to guide the client's behavior.
[0013] Step S2: Generation of excitation signal by the central server:
[0014] Before each training round begins, the server needs to estimate the data distribution state in the system and generate a set of "excitation signals" based on this state to guide the client's data selection. The specific steps are as follows:
[0015] Step S21: The server statistically analyzes the category distribution of the data uploaded by the client in the previous round, denoted as... As an estimate of the current system state, for each category The server calculates its proportion in the global distribution. This forms a global distribution estimate;
[0016] Step S22: Identify which categories of data are "scarce" or "oversaturated" globally, and generate activation vectors based on their scarcity level. Each element Reflection Category Scarcity or value;
[0017] Step S23: The server will use the scarcity vector Broadcast to all online clients as external information for their subsequent data selection strategies;
[0018] Step S3: The client selects data and trains the model based on the stimulus signal.
[0019] The client uses the stimulus signals provided by the server to autonomously select a subset of data for training, trains the model based on this subset, and then updates the model results to the server. The specific steps are as follows:
[0020] Step S31: The client receives the scarcity vector broadcast by the server. The system makes data selection decisions based on its own data conditions (local data distribution, data capacity, and training requirements) to maximize its own utility function and selects the subset of data to participate in this round of federated training. This serves as the data input for this round of training.
[0021] Step S32: The client selects its own subset of data. Complete local model training and upload the updated results to the server; simultaneously send the expected data subset distribution for the next round of client training.
[0022] Step S4, Global Model Update and Feedback:
[0023] The central server collects local model parameters or gradients uploaded by clients, aggregates them according to the set aggregation rules, and records the distribution of data subsets contributed by each client to estimate the new global data distribution state. The updated global model is then broadcast to each client to prepare for the next round of training. The specific steps are as follows:
[0024] Step S41: After each round of federated learning training is completed, the central server first receives the local model update results uploaded by each client, and assigns corresponding aggregation weights according to the size of the data subset uploaded by each client or the number of its effective samples. In this process, the model parameters of each client are weighted and summed according to the proportion of the samples they contribute, thereby forming new global model parameters.
[0025] Step S42: The server performs statistical analysis on the category distribution of all client-uploaded data in this round, updates the global data distribution estimate, and then calculates the scarcity score of each type of data based on the skewness of the current distribution, generating a new stimulus signal vector. This stimulus signal will be broadcast to the client before the start of the next training round to guide it to adjust its data selection strategy accordingly.
[0026] Compared with the prior art, the present invention has the following advantages:
[0027] 1. By guiding the client to select "scarce data" through a game theory mechanism, the global distribution of the system tends to be balanced, thus alleviating the training oscillation problem caused by non-IID.
[0028] 2. The client makes decisions independently locally without needing to upload complete data distribution or sample details, thus satisfying privacy constraints.
[0029] 3. The strategy is adaptively adjusted in each round based on the global state and local client information, making it suitable for dynamic data distribution scenarios.
[0030] 4. The game theory mechanism ensures that the client contributes to the global goal while maximizing its own utility. Attached Figure Description
[0031] Figure 1 This is a flowchart illustrating a data selection method for a game theory-based federated learning client. Detailed Implementation
[0032] The technical solution of the present invention will be further described below with reference to the accompanying drawings, but it is not limited thereto. Any modifications or equivalent substitutions to the technical solution of the present invention that do not depart from the spirit and scope of the technical solution of the present invention should be covered within the protection scope of the present invention.
[0033] This invention provides a data selection method for federated learning clients based on game theory, such as... Figure 1 As shown, the method includes the following steps:
[0034] Step S1, Game Theory Model Construction:
[0035] The client's data selection behavior is abstracted into a multi-agent non-cooperative game model. Before each round of federated training, the client independently selects a subset of data for training based on local observations and server feedback. This game model comprises the following components:
[0036] Participants: The set of clients currently active in federated learning. .
[0037] Client status information: for each client Have local dataset The corresponding data distribution is .
[0038] Action space: The client's policy space is all possible subsets of its local dataset, i.e., selecting a subset. This serves as the basis for this round of upload training.
[0039] Utility function: The client's benefit considers two factors: first, the contribution of the local data subset to the performance of the local model; and second, the support for the overall system's distributed equilibrium. This function is supported by scarcity incentive signals provided by the central server to guide client behavior.
[0040] By constructing the aforementioned game theory model, the system can guide individual behavior to serve the optimization goal of global data distribution while maintaining client autonomy.
[0041] The aim of this step is to formalize the client's data selection behavior in federated learning as a dynamic non-cooperative game process, so as to introduce game theory methods to optimize its strategy. Specific work includes:
[0042] Clearly define the participants in the game, namely all federated learning clients active in the current round;
[0043] Define the time structure of the game, that is, make a game decision before each round of federated training communication;
[0044] Establish client status information, including its local data distribution and data capacity;
[0045] Specify the client's actions, namely, selecting a subset of data from the local dataset for training;
[0046] Establish a utility function for each client, which considers both local data quality and contribution to global distribution balance, to encourage clients to make choices that are "beneficial to the system".
[0047] This step provides a game theory framework for strategy design, enabling subsequent strategies to be analyzed and controlled under distributed, dynamic, and information-constrained conditions.
[0048] Step S2: Generation of excitation signal by the central server:
[0049] Before each training round, the server needs to estimate the data distribution state in the system and generate a set of "stimulus signals" based on this state to guide the client's data selection. This step is performed by the central server, whose main task is to evaluate the global data distribution state and construct stimulus signals to guide the client's decision-making. This step includes the following operations:
[0050] Step S21: The server statistically analyzes the category distribution of the data uploaded by the client in the previous round, denoted as... As an estimate of the current system state, for each category The server calculates its proportion in the global distribution. This forms a global distribution estimate.
[0051] Step S22: Identify which categories of data are "scarce" or "oversaturated" globally, and generate activation vectors based on their scarcity level. Each element Reflection Category The scarcity or value of.
[0052] Step S23: The server will use the scarcity vector Broadcast to all online clients as external information for their subsequent data selection strategies.
[0053] The purpose of this step is to provide each client with feedback information related to the global state, guiding their data selection behavior to align with the overall system goals.
[0054] Step S3: The client selects data and trains the model based on the stimulus signal.
[0055] This step is performed locally by the client, which autonomously selects its training data subset using the stimulus signals provided by the server. Specifically, it includes the following steps:
[0056] Step S31: The client receives the scarcity vector broadcast by the server. The system makes data selection decisions based on its own data conditions (local data distribution, data capacity, and training requirements) to maximize its own utility function and selects the subset of data to participate in this round of federated training. This serves as the data input for this round of training.
[0057] During the optimization process, the client treats locally available data as a set of weighted items and maximizes the utility function while satisfying the upper and lower bounds of the categories that allow for selection of the subset size upper limit. This problem is essentially a combinatorial optimization problem of integer linear programming (NP-hard). Therefore, the client can use a greedy heuristic algorithm—first sorting the samples or categories by "unit utility" and then selecting sequentially up to the capacity limit, or further improving the quality of the solution by performing a local search (such as neighborhood swapping).
[0058] Step S32: The client selects its own subset of data. Complete local model training and upload the updated results to the server; at the same time, send the expected data subset distribution for the next round of client training.
[0059] Step S4, Global Model Update and Feedback:
[0060] This step is crucial for forming the game cycle, responsible for feeding back the game result of the current round to the next round, achieving long-term distribution self-optimization. The central server collects local model parameters or gradients uploaded by clients, aggregates them according to the set aggregation rules, and records the distribution of data subsets contributed by each client to estimate the new global data distribution state. The updated global model is then broadcast to all clients to prepare for the next round of training. The specific steps are as follows:
[0061] Step S41: After each round of federated learning training, the central server first receives the local model update results uploaded by each client. Based on the size of the data subset uploaded by each client or the number of valid samples, it assigns corresponding aggregation weights, typically using a weighted average method for global model updates, i.e., the typical FedAvg mechanism. During this process, the model parameters of each client are weighted and summed according to the proportion of samples they contribute, thus forming new global model parameters.
[0062] Step S42: The server statistically analyzes the category distribution of all client-uploaded data in this round, updates the global data distribution estimate, and then calculates the scarcity score of each type of data based on the skewness of the current distribution, generating a new stimulus signal vector. This stimulus signal will be broadcast to the client before the start of the next training round, guiding it to adjust its data selection strategy accordingly.
[0063] Through such a closed-loop process—the client selects a local data subset based on the scarcity signal, performs local training, and uploads the model update; the server updates the global model and data distribution accordingly and generates new incentive signals—a dynamic feedback mechanism is established between the client's behavior and the global objective, thus forming a continuous iterative, game-driven collaborative optimization process, enabling the system to gradually tend towards data distribution equilibrium and model stable convergence in multiple rounds of evolution.
Claims
1. A data selection method for federated learning clients based on game theory, characterized in that... The method includes the following steps: Step S1, Game Theory Model Construction: The client's data selection behavior is abstracted into a multi-agent non-cooperative game model. Before each round of federated training begins, the client independently selects the subset of data to be used for training in this round based on local observations and feedback information from the server. Step S2: Generation of excitation signal by the central server: Before each training round begins, the server needs to estimate the data distribution state in the system and generate a set of "excitation signals" based on this state to guide the client's data selection; Step S3: The client selects data and trains the model based on the stimulus signal. The client uses the stimulus signals provided by the server to autonomously select a subset of data for training, trains the model based on the subset, and updates the model results to the server. Step S4, Global Model Update and Feedback: The central server collects local model parameters or gradients uploaded by clients, aggregates them according to the set aggregation rules, records the distribution of data subsets contributed by each client to estimate the new global data distribution state, and broadcasts the updated global model to each client to prepare for the next round of training.
2. The data selection method for federated learning clients based on game theory according to claim 1, characterized in that... The game theory model comprises the following components: Participants: The set of clients currently active in federated learning. , Client status information: for each client Have local dataset The corresponding data distribution is ; Action space: The client's policy space is all possible subsets of its local dataset, i.e., selecting a subset. This serves as the basis for this round of upload training; Utility function: The client's benefit considers two factors: first, the contribution of the local data subset to the performance of the local model; and second, the support for the overall distribution balance of the system. This function is supported by the scarcity incentive signal provided by the central server to guide the client's behavior.
3. The data selection method for federated learning clients based on game theory according to claim 1, characterized in that... The specific steps of step S2 are as follows: Step S21: The server statistically analyzes the category distribution of the data uploaded by the client in the previous round, denoted as... As an estimate of the current system state, for each category The server calculates its proportion in the global distribution. This forms a global distribution estimate; Step S22: Identify which categories of data are "scarce" or "oversaturated" globally, and generate activation vectors based on their scarcity level. Each element Reflection Category Scarcity or value; Step S23: The server will use the scarcity vector Broadcast to all online clients as external information for their subsequent data selection strategies.
4. The data selection method for federated learning clients based on game theory according to claim 1, characterized in that... The specific steps of step S3 are as follows: Step S31: The client receives the scarcity vector broadcast by the server. The system combines local data distribution, data capacity, and training requirements to make data selection decisions in order to maximize its own utility function and select the data subset to participate in this round of federated training. This serves as the data input for this round of training. Step S32: The client selects its own subset of data. Complete local model training and upload the updated results to the server; at the same time, send the expected data subset distribution for the next round of client training.
5. The data selection method for federated learning clients based on game theory according to claim 1, characterized in that... The specific steps of step S4 are as follows: Step S41: After each round of federated learning training is completed, the central server first receives the local model update results uploaded by each client, and assigns corresponding aggregation weights according to the size of the data subset uploaded by each client or the number of its effective samples. In this process, the model parameters of each client are weighted and summed according to the proportion of the samples they contribute, thereby forming new global model parameters. Step S42: The server performs statistical analysis on the category distribution of all client-uploaded data in this round, updates the global data distribution estimate, and then calculates the scarcity score of each type of data based on the skewness of the current distribution, generating a new stimulus signal vector. This stimulus signal will be broadcast to the client before the start of the next training round to guide it to adjust its data selection strategy accordingly.