Incremental Learning-Based Radio Frequency Fingerprint Identification via Exemplar Replay

With the development of the Internet of Things, the number of wireless communication devices is rapidly increasing, which has brought security and privacy issues into great attention. Radio frequency fingerprint (RFF) identification is a key technique for ensuring communication security at the physical layer. Deep learning (DL)-based methods have achieved wonderful performances in RFF identification. However they usually suffer from catastrophic forgetting when the new data differs from the former data. Therefore, incremental learning (IL) is viewed as a viable solution to address the previously mentioned issues. In this paper, we propose an IL-based RFF identification method via exemplar replay, which combines knowledge distillation and exemplar replay. Experimental results indicate that, upon the arrival of new data streams, the proposed method exhibits higher improvement in accuracy compared to the performance of feature extraction, ultimately reaching 96.21 %. Furthermore, the size of the exemplar set is merely 2.5 % of the total training set.