Optimized Satellite Participation in Federated Learning over LEO Constellation Networks

This paper presents an intelligent device selection framework for Federated Learning (FL) in Low Earth Orbit (LEO) satellite constellation networks. In particular, the high mobility, intermittent connectivity and long communication delays in LEO satellite networks significantly impact FL convergence and model performance. To address these challenges, we formulate a device selection optimization problem that determines the optimal subset of LEO satellites to participate in global model aggregation at each round. The objective is to minimize the average loss while satisfying latency constraints. To solve the formulated problem, we develop a deep reinforcement learning (DRL)-based solution that enables adaptive satellite selection considering the stochastic nature of satellite connectivity and channel variability. The proposed framework selects participating satellites dynamically based on their communication latency and computational capabilities to improve the overall efficiency of the FL process. Simulation results show that the proposed approach accelerates convergence compared to conventional synchronous FL while maintaining model accuracy.