Research On Optimization Of Multi-user Federated Learning System Based On Game Theory

To overcome the challenges of security and privacy problems in centralized machine learning,federated learning(FL)enables collaborative learning between devices without sharing the local training data with any third-party entity.However,FL trains a single global model and requires all participating devices to communicate frequently with a central server to transmit model parameters.When the training data are not independent and identically distributed(IID)across different devices,it is difficult for the central server to train a model that will work well for all user devices.Additionally,some devices are unable to participate in FL because of energy limitations or a significant transmission delay.Moreover,as the behavior of each user in the FL system is not controlled by the server,it is necessary to consider users’ willingness to ensure that a sufficient number of users participate in the training.To solve these problems,the design scheme of incentive mechanism and the optimization of network topology structure in multi-user FL system are studied by game theory in this paper.The main research contents include:(1)To solve the user incentive issues in clustered federated learning(CFL),the participating-willingness of the user in the real world is taken into account and an effective incentive mechanism is designed to enhance the performance of users’ local models in the context of data heterogeneity.First,the utility functions of the central server and users are designed respectively.The model distance and local accuracy are used to measure a user’s contribution,and the communication cost and computation cost of users are considered jointly.The overall performance of the multiple intracluster center models on the server is evaluated by the weighted average of the local absolute accuracy of participating users,which is used to design the reward function of the server.Next,we formulate the optimization problems as a two-stage Stackelberg game with the knowledge of the utility functions of both parties,where the optimization problem of the server is defined as maximizing the expectation of multiple rounds of accumulated utility and the deep reference learning(DRL)approach is utilized to find an optimal strategy for the server.The incentive mechanism designed in this paper can effectively motivate rational users to participate in the intracluster model training.When there are low-performing devices in the system,the server can balance the number of users and the training efficiency to encourage high-quality users to participate in the model training.(2)Based on the idea of cooperative federated learning,a new hierarchical network structure is built to address the issue of low user data utilization under the FL star network architecture,which improves the accuracy of the global model by increasing the number of users through the cooperation among them.A two-layer game is proposed to analyze the formulation process of the network and a stable network topology will be formed ultimately.First,a set of relay devices are selected by the server by managing the reward rate for users,and the interaction between the server and the users is modeled as a Stackelberg game.Next,the connections between the relay devices and slave devices are determined by a matching game.Preference functions for both parties are designed to enable them to rank each other according to several factors such as the amount of data and energy consumed.The relay devices and slave devices are matched according to their preferences until no devices deviate unilaterally and a stable connection will be obtained.Experiments show that the hierarchical network architecture can effectively increase the utilization rate of training data in the system,and the test accuracy of the global model can be improved significantly when data samples across all user devices are IID.