Generalized Category Discovery in Radio Frequency Fingerprint Identification: An End-to-End Approach with Signal-Augmented Entropy Regularization

The identification and clustering of new wireless devices outside known categories are crucial for radio frequency fingerprint identification (RFFI), particularly important for device authentication and intrusion detection. However, most existing RFFI schemes have low robustness to environmental changes, limited scalability, and poor generalization across different scenarios. To overcome these shortcomings, we propose an end-to-end RFFI approach with signal-augmented entropy regularization (E2E-SAER). Specifically, we first construct a feature extractor based on a multi-block mixture of experts to capture condition-specific and radio frequency fingerprint-related features under inherent channel/signal diversity. We then propose an end-to-end approach for fingerprint classification/clustering to further optimize the feature extractor for more effective RFFI. Since the end-to-end training may result in model overfitting to labeled data and the biased predictions towards known devices, our E2E-SAER enhances the prototype classifier with a signal-augmented entropy regularization to achieve more uniformly distributed predictions between known and unknown categories for effective known class identification and unknown class clustering. We conduct extensive experiments on seven open-source radio frequency fingerprint datasets with six benchmarks and demonstrate that our proposed E2E-SAER significantly outperforms existing algorithms.