SEI-DSR: Specific Emitter Identification via Dynamic Sparse Attention and Multidimensional Feature Reconstruction

Specific Emitter Identification (SEI) is a physicallayer authentication technique based on radio frequency fingerprinting. Although deep learning has made significant progress in the SEI domain in recent years, existing methods still face several challenges: traditional global attention mechanisms indiscriminately process all features, amplifying noise interference alongside useful signals, thereby increasing susceptibility to noise; additionally, current methods underutilize frequency-domain information. To tackle these challenges, this paper introduces the SEI-DSR method. This approach innovatively integrates dynamic sparse attention, multidimensional redundancy suppression, and frequency-domain feature enhancement techniques. Specifically, it employs Bi-Level Routing Attention (BRA) to first dynamically filter out redundant regions, followed by fine-grained attention computation within key regions, significantly enhancing the focus of model on critical features. It also incorporates Spatial-Channel Convolution (SCConv) to jointly optimize feature redundancy in both spatial and channel dimensions, further improving the feature of model selection capability. Moreover, a module for frequency-domain enhancement is incorporated to improve performance in environments with a low signal-to-noise ratio (SNR). Experimental results demonstrate that the proposed method outperforms existing approaches in identification accuracy, with particularly standout performance in low-SNR electromagnetic environments.