Dynamic Games in Federated Learning Training Service Market

Abstract

With the great success of deep learning and increasingly powerful mobile devices, federated learning gains growing attentions from both academia and industry. It provides high-quality on-device model training services while preserves data privacy. In this paper, we consider a federated learning training service market which consists of model owners as consumers and mobile device groups as providers. A two-layer dynamic game is formulated to analyze the dynamics of this market. In particular, the service selection processes of model owners are modeled as a lower-level evolutionary game while the pricing strategies of mobile device groups are modeled as a higher-level differential game. The solutions of the dynamic games, i.e., dynamic equilibrium are analyzed theoretically and verified via extensive numerical evaluations.