Federated Learning over LEO Satellite

The rapid development of low earth orbit (LEO) satellite communication has driven the deployment of artificial intelligence (AI) in space, providing various intelligent services like real-time disaster navigation, global pandemic spread detection, etc. To this end, a space-ground communication based satellite federated learning framework is proposed in this work. In the framework, a LEO satellite works as a server and collaboratively trains an AI model with multiple ground devices distributed in a large remote area, where there are no ground access points (APs) due to the expensive cost. To overcome the communication bottleneck and address the new challenge arising from the satellite-ground communication channel model, we propose an over-the-air computation (AirComp) based FL scheme and take into account the influence of both uplink and downlink communications. Then, the convergence bound is analyzed in the proposed scheme, where the optimality gap in each training iteration depends on two individual factors, i.e., the downlink and uplink errors. Accordingly, two sub-problems are formulated to minimize the two kinds of errors. The downlink sub-problem is convex and can be addressed by the well-known CVX toolbox. For the non-convex uplink problem, inspired by the successive convex approximation (SCA) algorithm, we propose a SCA-based bounded perturbation (BSBP) algorithm. Extensive numerical results show that the proposed algorithm can significantly enhance the FL performance with low complexity.