Mobile Device Training Strategies in Federated Learning: An Evolutionary Game Approach

Abstract

With the tremendous success of machine learning and increasingly powerful mobile devices, federated learning has gained growing attention from both academia and industry. It capitalizes on vast number of distributed data to support machine learning based applications while maintaining data privacy. In this paper, we consider a federated learning system, in which the mobile devices allocate their data and computation resources among the machine learning applications, i.e., model owners. Specifically, we formulate an evolutionary game mode for the mobile devices with bounded rationality to adapt their training strategies aiming to maximize the device's individual utility. The uniqueness and stability of the equilibrium of the game are analysed theoretically. Besides, Extensive experiments are conducted to determine the functions fitted for the accuracy and energy consumption metrics.