Boosting Semi-Supervised Specific Emitter Identification with Curriculum Triple Consistency Regularization

Specific Emitter Identification, a critical technique for physical layer authentication in the Internet of Things, extracts Radio Frequency Fingerprints from signals affected by hardware impairment to discern between devices. The soaring development of the radio industry and the exorbitant expenses associated with signal labeling collectively necessitate the adoption of semi-supervised learning. This study proposes an approach employing curriculum triple consistency regularization to enhance Semi-Supervised Specific Emitter Identification. We incorporate constraints on weak-strong, branch, and semantic consistency, while implementing curriculum thresholds that dynamically filter pseudo-labels during training. Our approach achieves significant enhancements in classification and clustering performance, especially when faced with challenging training scenarios, while maintaining a lightweight codebase and low time and space complexity. Taking the 20-class ADS-B dataset with a 5% labeling ratio as an example, our approach achieved 77.80% accuracy and a Silhouette Coefficient of 0.272106, representing improvements of 13.00% and 0.138745, respectively.