CTAM-ResNeSt: A Specific Emitter Identification Method Driven by Cross-Dimensional Triplet Attention and Multi-Scale Feature Fusion

Amid the rapid evolution of wireless communication systems, specific emitter identification (SEI) technology within physical layer security encounters emerging technical challenges. To address the shortcomings of existing deep learning methods in terms of insufficient cross-dimensional feature interaction and limited local feature extraction capability, this paper proposes a SEI method driven by cross-dimensional triplet attention and multi-scale feature fusion, termed CTAM-ResNeSt. This approach utilizes a lightweight ResNeSt-50 network as its core architecture, innovatively embedding a lightweight Cross-Dimensional Convolutional Triplet Attention Module (CTAM). Specifically, the CTAM module simultaneously computes interaction weights across the channel, height, and width dimensions, establishing a dynamic cross-dimensional feature interaction mechanism. Meanwhile, the enhanced ResNeSt-50 network employs a multi-branch structure with cardinal grouping and split subgroups to achieve hierarchical feature extraction and attention-weighted fusion. Experiment results demonstrate that, compared to mainstream baseline networks, the proposed method exhibits significant advantages in recognition accuracy and robustness, outperforming existing approaches across multiple benchmark datasets.