A Novel Approach to Continuous Security Authentication in the Internet of Things Based on RF Fingerprints

Security authentication is of paramount importance in the Internet of Things (IoT) applications. Radio frequency fingerprints, which represent unique physical layer characteristics of emitters, offer a promising authentication approach without imposing additional overhead, and have garnered significant attention in the realm of IoT security authentication. Given the open nature of IoT environments, new emitters continually seek to join the network. However, the computational demands associated with retraining the identification network are exorbitant. Fine-tuning methods may introduce challenges such as the forgetting of historical knowledge and susceptibility to incorrect labels. This paper introduces a novel method for continuous security authentication in IoT based on RF fingerprints. It leverages base class knowledge distillation and data replay to mitigate the forgetting of historical RF fingerprints knowledge. Furthermore, it designs a sample feature prototype evaluation algorithm to selectively replay samples, thereby further reducing the number of replay samples required. Additionally, a loss function that combines self-supervised learning and robust rank statistics is proposed to mitigate the impact of incorrect labels associated with newly added emitters. Experimental results demonstrate that even after adding 8 emitters, the proposed algorithm maintains an accuracy of 96.22%, which is only 3.62% lower than the performance upper bound.