In-Orbit Aggregator-Empowered Federated Learning Framework for Satellite and Terrestrial-Integrated Networks

In this paper, we introduce a federated learning (FL) framework tailored for a satellite and terrestrial-integrated network (STIN), which employs a semi-asynchronous FL algorithm and in-orbit aggregations (IOA) to mitigate the straggler issue and enhance energy efficiency. Our goal is to optimize IOA-aware routing to enable energy-efficient model aggregation with uncertain ground stations (GSs) in terms of upload-ready timing. To this end, we utilize a time-expanded directed graph (TEDG) to effectively account for the network's connectivity and energy demands. Furthermore, we propose a predictive algorithm to cope with the uncertainty of GSs. A preliminary result demonstrates the robustness of our approach even under inaccurate predictions, achieving a marginal gap of 2% of the cost compared to the optimal scheme.