Physical Layer Authentication Enhanced Federated Learning for UAV with Intrinsic Security

Unmanned aerial vehicles (UAVs) can be utilized effectively as airborne base stations, offering wireless communication and federated learning (FL) services for terrestrial edge devices (EDs). FL enables EDs to collaboratively train a global model for specific tasks without sharing their local raw data. However, due to the unreliable wireless links and dynamic topologies, potential malicious nodes may attempt to masquerade as legitimate nodes to pose various potential threats (e.g., inference, poisoning, and backdoor attacks) to undermine the trustworthiness of the intermediate model parameters. Although existing cryptographic authentication protocols focus on the data-level security, security level can be further enhanced at the physical layer level. In this paper, we design FedPLA, a UAV-aided federated learning framework enhanced by physical layer authentication, which utilizes SNR difference for lightweight authentication to ensure FL intrinsic security. Subsequently, we formulate a multi-step decision problem for joint UAV trajectory and resource allocation, aiming to minimize the latency and energy costs while maximizing secure edge device number with intrusion-proof and unjustly accused-proof guaranty. To efficiently deduce strategies while avoiding potential dangerous policies, we develop an LSTM-enhanced safe deep reinforcement learning algorithm (LSTM-SDRL) for real-time strategy making. Furthermore, extensive simulations demonstrate the effectiveness of the proposed LSTM-SDRL.