Improve Global Generalization for Personalized Federated Learning Within a Stackelberg Game

Abstract

Maintaining the personalization of local models under unique local data has driven the development of various personalized federated learning approaches. However, these methods often neglect the inference and generalization capabilities of the aggregated global model. Ensuring optimal local personalization with enhanced global generalization remains a significant challenge. In this work, we introduce personalized federated learning within a Stackelberg strategy, a novel personalized federated learning framework aimed at improving global generalization in personalized federated learning. We formulate federated learning as a Stackelberg game and leverage the soft actor-critic reinforcement learning method to explore the optimal global model, considering the dynamic changes in model parameters and feature outputs.