Cross-Attention Transformer for Channel-Robust Radio Frequency Fingerprint Identification

Deep learning-based Radio Frequency Fingerprint Identification (RFFI) has emerged as a promising technique owing to its impressive performance in wireless emitter identification. In real-world scenarios, variations in wireless channel states during training and testing lead to drifts in data distribution, significantly reducing identification accuracy. To address this problem, a cross-attention transformer for channel-robust RFFI is proposed in this work. First, a fingerprint feature separation method based on similarity measurement is proposed to effectively preserve hardware features while weakening channel features by exploiting their different characteristics. Second, the Segmented Logarithmic Spectra Embedding (SLSE) module and the Multi-Head Cross-Attention (MHCA) mechanism are designed to construct the identification network. The designed network focuses more attention on the essential characteristics of the emitter, significantly enhancing its robustness to channel variations. Experiments on a public WiFi signal dataset show that the proposed method can effectively reduce the impact of channel variations on the Radio Frequency Fingerprint (RFF) features, achieving an identification accuracy of 99.72%.