Specific Emitter Identification with Few-Shot via Deep Networks Based on Time and Frequency Domain with Channel, Spatial, and Self-Attention Mechanisms

Specific emitter identification (SEI) is a highly active research area in physical layer security. In this paper, we propose a SEI scheme based on time-frequency domain channel, spatial, and self-attention mechanisms (TF-CSS) for deep networks with few-shot learning. The scheme first uses an asymmetric masked auto-encoder (AMAE) with attention mechanisms for unsupervised learning, then removes the decoder and adds a linear layer as a classifier, and finally fine-tunes the whole network to achieve effective recognition. The scheme improves the feature representation and identification performance of complex-value neural network (CVNN)-based AMAE by adding channel, spatial, and self-attention mechanisms in the time-frequency domain, respectively. Experimental results show that this scheme outperforms the recognition accuracy of contrastive learning and other MAE/AMAE-based methods in 30 classes of LoRa baseband signal transmitter recognition with different few-shot scenarios and observation lengths.