面向星地协同网络的联邦边缘学习方法研究

Federated Edge Learning Approaches for Satellite-Terrestrial Cooperative Networks:A Study

目前地面蜂窝网络无法实现全球广域覆盖,同时抗毁性不高。低轨卫星网络的加入,弥补了地面蜂窝网络的不足。二者协同成为星地协同网络后可以打破两种网络之间互相独立的现状并结合它们的优点,是未来的发展趋势。在对星地协同网络中海量设备产生的大量数据进行分析处理时,易出现隐私信息泄露和数据孤岛问题。为解决上述问题,提出一种包含分层聚合和用户关联策略的联邦边缘学习算法。该算法通过分层聚合来适配星地协同网络,并通过用户关联策略实现时延和模型精度的联合优化。仿真表明,所提出的算法可以使协同网络中进行联邦边缘学习的全过程时延较低,同时可以得到较高的模型测试精度。

Current terrestrial cellular networks cannot achieve global wide-area coverage and lack robustness against destruction.The integration of low Earth orbit satellite networks compensates for these deficiencies.The collaboration of these two,forming satellite-terrestrial cooperative networks,breaks the independence of both systems and combines their advantages,signaling a future trend.When analyzing and processing massive amounts of data generated by numerous devices in satellite-terrestrial cooperative networks,issues of privacy leakage and data silos often arise.To address these challenges,a federated edge learning algorithm encompassing hierarchical aggregation and user association strategies is proposed.This algorithm adapts to satellite-terrestrial cooperative networks through hierarchical aggregation and achieves joint optimization of latency and model accuracy through user association strategies.Simulations show that the proposed algorithm facilitates federated edge learning in cooperative networks with lower overall latency,while also achieving higher model test accuracy.