Staleness-Control Semi-Asynchronous Satellite Federated Learning via Flexible Aggregation

Satellite Federated Learning (SFL) has emerged as a viable approach for deploying distributed machine learning in Low Earth Orbit (LEO) satellite networks, enabling timely processing of space-acquired data without the need to transmit raw data back to ground. However, in the constellations with ground-satellite links, intermittent visibility windows and Non-IID data distributions significantly exacerbate both the occurrence and negative impacts of staleness in semi-asynchronous SFL. Specifically, outdated local models uploaded by satellites may compromise convergence stability. To address these challenges, we propose a staleness-control semi-asynchronous SFL framework. By leveraging Lyapunov optimization, it maintains staleness below a predefined threshold while dynamically adjusting the number of fresh models aggregated per round, thereby maximizing communication window utilization and enhancing training efficiency. To further mitigate the impact of staleness, the staleness-aware weights are applied during aggregation. Additionally, a gradient compensation strategy is incorporated, which leverages historical models from non-updated satellites to mitigate the impact of Non-IID data. Experimental results on MNIST and CIFAR-10 show that our method speeds up training by 1.25× to 1.95× over baseline approaches.