Adaptive Satellite-to-Device Association Based Asynchronous Federated Edge Learning in STIN

Satellite-Terrestrial Integrated Networks (STIN) has attracted attention because of its advantages in low latency, wide coverage, and strong robustness. The analysis of data collected in STIN can lead to privacy leakage and communication overhead. In addition, the dynamic topology of Low Earth Orbit (LEO) satellites and the heterogeneous device computational capabilities make machine learning inefficient. In this paper, an Asynchronous Federated Edge Learning (AFEL) algorithm is proposed to tackle these challenges in STIN. Specifically, devices conduct local training, while LEO satellites and a Geostationary Earth Orbit (GEO) satellite perform the edge-level and global aggregation, respectively. The impact of AFEL algorithm parameters on convergence is also analyzed, and appropriate adjustments facilitate more stable convergence. To satisfy diverse task demands in model accuracy and overall delay, we incorporate an Adaptive Satellite-to-Device Association (ASDA) strategy into AFEL under the real visibility in STIN, enabling joint optimization of overall delay and model accuracy. Extensive simulations demonstrate that AFEL achieves high model accuracy under suitable parameters. Moreover, with the ASDA strategy, the AFEL can flexibly balance model accuracy and overall delay based on task requirements. This paper introduces an asynchronous paradigm for federated learning in STIN and explicitly account for real visibility relationships to improve both latency and accuracy.