Covert Air–Ground Federated Learning With Pruning: Enabling by STAR-RIS

Abstract

Federated learning (FL) can safeguard the data privacy for air-ground networks, while simultaneously transmitting and reflecting reconfigurable intelligent surface (STAR-RIS) can help further enhance its channel quality. However, the model exchanges in FL may be detected by wardens, and air-ground links increase this threat. In this paper, we investigate a STAR-RIS aided air-ground covert FL scheme to hide the presence of uploading, where the unmanned aerial vehicle (UAV) acts as an aerial server to collect local models from the distributed devices. To alleviate the communication burden, each device prunes the local model before transmitting. The minimum detection error probability at the warden is derived with the optimal detection threshold, indicating the worst case for the legitimate covert FL. Then, we analyze the convergence of the proposed FL scheme and provide an upper bound on the expected convergence gap of global loss function. To improve the learning performance, the convergence gap is minimized subject to the covertness and latency constraints. Specifically, the closed-forms of optimal pruning ratio and transmit power are derived, while the phase shifts of STAR-RIS and UAV position can be optimized by semidefinite relaxation and successive convex approximation, respectively. Finally, the overall problem can be solved via an alternative algorithm. Numerical results validate that the proposed covert air-ground FL scheme can achieve excellent test accuracy and convergence performance while guaranteeing the covertness assisted by STAR-RIS.