Optimizing Contact-Based Decentralized Satellite Federated Learning

The integration of LEO satellite onboard processing with federated learning has propelled a promising paradigm: satellite federated learning (SFL), by empowering onboard machine learning (ML) to provide various intelligent services. Especially, the decentralized satellite federated learning (DSFL), where each LEO satellite serving as a client exploits one-hop inter-satellite links (ISLs) to exchange local models, could alleviate the reliance on the centralized ground server, and reduce multi-hop model transmissions. In a DSFL system, the scheduling of model training and transmission significantly affects the learning performance. Thus it is crucial yet challenging to design an efficient scheduling strategy for the dynamic LEO satellite networks with heterogeneous onboard datasets. In this paper, we propose a contact-based DSFL framework, where each contact between a pair of satellites is regarded as a collaboration opportunity to exchange local models. Under this framework, we formulate a problem of optimizing the scheduling strategy with the aim of maximizing the DSFL model accuracy. To solve this problem, we design a Double deep <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$Q$</tex> learning (DDQN) based scheduling strategy, which schedules the local model training and transmission upon each contact by leveraging the instantaneous environmental information, such as model accuracy, the neighbor's model accuracy, and available training time. Simulation results demonstrate the effectiveness and efficiency of the proposed DDQN-based scheduling strategy over three baselines.