Detecting Model Poisoning Attacks Via Dummy Symbol Insertion for Secure Over-the-Air Federated Learning

With the rapid advancement of intelligent IoT services, the need for high-quality and efficient wireless communication has become increasingly critical for information exchanges and collaborations among intelligent edge devices such as autonomous mobile robots (AMRs), robots in smart factories, and surveillance cameras, etc., Federated Learning (FL) is a promising distributed learning framework by exploring the computation capability on edge devices while protecting data privacy of users. On the other hand, Over-the-Air (OTA) computation techniques is a new paradigm of integrated computation and communication. OTA based FL (OTA-FL) can enhance spectrum efficiency by leveraging the superposition properties of wireless channels for model aggregations. The implementation of OTA-FL on edge devices can further protect user data privacy for clients, improve the throughput of inter-device communications, and enhance the distributed learning performance. However, despite these advantages, the convergence of OTA-FL systems remain vulnerable to model poisoning attacks, presenting significant challenges. In this study, we address these vulnerabilities by proposing a two-phase detection mechanism that secures OTA-FL system while preserving high spectral efficiency. Our simulation results demonstrate that under varying conditions, our approach can make distributed learning more secure and communication-efficient.