Energy-Aware Edge Association for Cluster-Based Personalized Federated Learning

Federated Learning (FL) over wireless network enables data-conscious services by leveraging the ubiquitous intelligence at edge for privacy-preserving model training. As proliferation of context-aware services, diversified personal preferences cause disagreeing conditional distributions among user data, which leads to poor inference performance. In this sense, clustered federated learning is proposed group devices with similar preference and provide each cluster a personalized model. This calls innovative design in association that involves clustering also resource management optimization. We formulate an accuracy-cost trade-off optimization problem jointly considering accuracy, communication allocation energy consumption. To comply parameter encryption techniques FL, we propose iterative solution procedure employs deep reinforcement based approach cloud server association. The reward function consists minimized consumption base station averaged accuracy all users. Under our solution, multiple are fully exploited realize cost efficient without any prior knowledge on parameters. Simulation results show strategy outperforms existing strategies achieving accurate low cost.