Distributed Federated Learning in Satellite Constellations: A Framework for in Orbit ML

In the rapidly evolving field of space operations, especially within satellite constellations, the traditional method of collecting data from multiple satellites into a central hub for machine learning tasks is becoming too costly and impractical. This centralization not only requires significant resources and communication overhead but also poses serious privacy risks, assuming that each satellite holds private data for model training purposes. As the future of space communication and observation shifts towards using more satellite constellations, there is a growing need for a scalable and resource-efficient approach to distributed machine learning in space. Our new framework, SatFL, tackles these challenges by enabling each satellite to train a local machine learning model on its own data. These individual models are then combined through a process of iterative refinement across the constellation to form a unified global model. This framework avoids the need for direct data sharing between satellites, which preserves data privacy, reduces bandwidth demands, and conserves the limited energy resources of each node. SatFL is specifically designed for the unique conditions faced by space-based nodes, functioning autonomously without relying on any central authority or leader - vital due to the intermittent connection of satellites to ground stations and other space assets. By decentralizing the learning process, SatFL enhances the resilience and scalability of machine learning applications in space. It ensures optimal use of the limited computational and energy resources available to each satellite, addressing one of the most significant challenges in space technology. Our experimental results clearly demonstrate that distributed federated learning (FL) is not only feasible but also offers the scalability required for future complex space missions where the presence of a central authority is unlikely.