Few-Shot Specific Emitter Identification Based on Dual Dynamic Universum Contrastive Learning

Abstract

Specific emitter identification (SEI) is a critical technique for physical-layer security and is widely applied in non-cooperative communications and spectrum management. Current deep neural network approaches for SEI exhibit a strong dependence on labeled data. Moreover, due to subtle radio-frequency fingerprints, they suffer from uninformative negative samples and blurred inter-class boundaries in few-shot regimes. To address these issues, we innovatively apply Universum learning to solve the SEI problem, proposing a few-shot framework termed Dual Dynamic Universum Contrastive Learning (DDU-SEI). The proposed framework enables self-supervised pre-training via physics-aware dual-domain augmentations, adaptive anchor-specific Universum generation, and multi-objective joint optimization. Experiments on the WiFi and ADS-B datasets demonstrate that DDU-SEI consistently outperforms representative self-supervised baselines, especially under the extremely low-shot scenario.