Cross-attention relation network based on metric learning for few-shot specific emitter identification

Specific Emitter Identification (SEI) based on radio frequency fingerprinting has emerged as a promising physical-layer authentication technique in wireless communication security. Although deep learning has demonstrated powerful capabilities in extracting fingerprint features, existing learning paradigms heavily rely on precise label information and large amounts of training data, which poses significant limitations in real-world scenarios with few-shot conditions. Inspired by metric learning, we propose a few-shot SEI method based on cross-attention relation network (CARN). To tackle the challenge of poor generalization due to limited data, we introduce a relation network for computing nonlinear similarity metric. By leveraging contrastive learning with meta-task training, the model rapidly adapts to few-shot scenarios. For cross-channel emitter identification, we incorporate a cross-attention mechanism, enabling the model to focus on key regions in the time-frequency spectrogram while preserving channel information. Extensive experiments demonstrate that, compared to state-of-the-art methods, CARN achieves outstanding recognition performance and remarkable stability on the public LoRa dataset. Specifically, CARN improves unseen emitter identification by 37.5 % in the 5-shot scenario.