Efficient Space-Ground Federated Learning Based on Compressed Sensing

Satellite communication (SatCom) networks based on low Earth orbit (LEO) satellites play an important role in the sixth-generation mobile communication network (6G) and Terrestrial-Aerial-Space network. However, traditional SatCom faces the following challenges: 1) The lack of computing power of satellite nodes. 2) Limitations of communication bandwidth. 3) The risk of node data being leaked. To tackle the challenges, in this paper, we propose an efficient space-ground federated learning (FL) algorithm based on compressed sensing (CS). The proposed framework improves upon the existing frameworks by allowing terminal nodes to send quantized data to the central server after compression with a random measurement matrix. Based on that, the central server updates the global model parameters utilizing the result of sparse reconstruction obtain by Bilinear Adaptive Generalized Vector Approximate Message Passing (BAd-GVAMP). Experiments on both independent and identically distributed (IID) datasets and non-IID datasets show that the proposed framework is able to reach high training accuracy while achieving substantial compression of model parameters and communication delay. The proposed framework enables high-efficiency and high-precision space-ground FL by fully utilizing the computing power of terminal nodes, tackling the communication bottleneck problem of SatCom, and protecting node data from leakage.