Toward Federated Learning With Byzantine and Inactive Users: A Game Theory Approach

Abstract

Federated learning (FL) can guarantee privacy by allowing local users only upload their training models to central server (CS). However, the existence of Byzantine or inactive users may cause model corruption or inactively participation in FL. In this paper, a game theory based detection and incentive method is designed for Byzantine and inactive users. Specifically, a differential aggregate gradient descent (DAGD) algorithm is adopted to improve the stability and fasten the convergence. Then the loss function is modified by considering Byzantine and inactive users. For Byzantine users, a random Euclidean distance (RED) voting method is designed to identify Byzantine users, and after identification, Byzantine users are motivated by the game theory. For inactive users, when inactive users are detected by contribution, Nash equilibrium with a mixed strategy is used to solve the problem of inactive users' participation in FL. Extensive experimental results show that Byzantine users and inactive users can be detected and motivated by the algorithm. Meanwhile, compared with other methods, the accuracy of the optimized model is improved with reduced training time.