Hierarchical Federated Edge Learning over Space-Air-Ground Integrated Networks

Federated learning (FL), as a key application scenario to support distributed artificial intelligence (AI) services, allows an amount of edge devices to train a global model in tandem. In addition, space-air-ground (SAG) integrated networks are envisioned to provide AI services for remote areas in the next generation of wireless communication system, which provides new opportunities for the integration of SAG and FL. Whereas, the performance of the SAG-FL is limited by significant system latency, high energy consumption, and slow convergence rate. In this paper, we consider a SAG hierarchical FL network, where a low-earth-orbit (LEO) satellite serves as the cloud server and multiple unmanned aerial vehicles (UAVs) act as the edge nodes covering multiple cells of edge devices. Specifically, we provide the convergence analysis of the SAG-FL algorithm, and then alternatively optimize the device scheduling policy and UAVs' trajectory to minimize the system delay and energy consumption between two global aggregations. Simulation results illustrate that the proposed dynamic device scheduling policy is much more time-saving and energy-efficient than the static one.