Flexible Synchronous Federated Learning Approach for LEO Satellite Constellation Networks

This paper presents a flexible synchronous federated learning (FlexSync-FL) approach for Low Earth Orbit (LEO) satellite constellation networks, where LEO satellites train local models and conduct a collaborative global model at the Network Operations Center (NOC). Unlike the standard synchronous FL, FlexSync-FL employs a dual-trigger synchronization mechanism that initiates global model aggregation either upon receiving updates from all clients (satellites) or after a predefined maximum interval time has elapsed. Furthermore, FlexSync-FL leverages inter-satellite links (ISLs) to facilitate forwarding local models among satellites, especially for those without direct visibility to ground gateway stations (GWs). In particular, FlexSync-FL aims to mitigate the impact of long latency and intermittent connectivity, inherent in satellite networks, on the FL process. The effectiveness of the proposed FlexSync-FL framework is demonstrated through simulations that employ Long ShortTerm Memory (LSTM) networks to train local models at each LEO satellite for traffic forecasting using real-world aeronautical datasets.