Decentralized Satellite Federated Learning via Intra- and Inter-Orbit Communications

Satellite federated learning (SFL) emerges as $a$ promising approach to exploit computing and data resources within low Earth orbit (LEO) satellite constellations for supporting intelligent applications while maintaining privacy protection. Most of the existing works of SFL consider $a$ ground station as the central server to aggregate model pa-rameters transmitted from satellites. However, the intermittent connection between satellites and ground stations leads to overlong convergence time. In this paper, we propose a novel decentralized SFL in the Walker-Delta constellation, where satellites communicate with nearby satellites in the same or different orbital planes through inter-satellite links (ISLs). We formulate a mixed combinatorial optimization problem that involves the joint optimization of power control, bandwidth allocation, and routing selection of each satellite to minimize the energy consumption in SFL. A joint resource allocation and routing selection (JRARS) algorithm is subsequently developed to solve the proposed optimization problem. Results show that our proposed decentralized SFL framework outperforms other SFL frameworks in terms of convergence time and accuracy. We also show that our proposed JRARS algorithm consumes much lower energy consumption than the baseline algorithms when the model converges.