Dual-Prior Generative Augmentation for Few-Shot Specific Emitter Identification

Abstract

Specific emitter identification (SEI) provides physical-layer device authentication by exploiting hardwareinduced radio frequency fingerprints (RFFs). However, deep learning-based SEI methods face severe performance degradation when labeled samples are scarce due to high collection costs and dynamic deployment conditions. To address this few-shot challenge, we propose a dual-prior generative augmentation (DPGA) framework that jointly enhances data diversity and classification discrimination. The proposed method first converts I/Q signals into contour stella images (CSIs) for visual representation. A dual-prior generative adversarial network (DP-GAN) is then designed with heterogeneous CNN-Transformer feature extraction to produce high-quality synthetic samples, incorporating a mode-seeking diversity regularization to enlarge the coverage of the generated distribution. Finally, a cosine prototype classifier aggregates both original and synthetic samples and performs emitter classification based on prototype similarity in the feature space. Experimental results on a real-world RF fingerprint dataset demonstrate that DPGA consistently outperforms existing few-shot SEI methods.