RF Fingerprinting Based on DAFFM in Low SNR

The rapid growth of the Internet of Things (IoT) has led to a surge in the number of devices accessing wireless networks, raising an urgent need for secure authentication. Radio Frequency Fingerprinting (RFF) utilizes unique signal features from hardware imperfections to identify devices. It has emerged as a promising physical-layer solution due to its passive operation and the elimination of key management requirements. However, under complex and noisy channels, traditional RFF methods suffer from performance degradation. To address this issue, we propose a robust RFF method based on a Dual-path Attention Feature Fusion Module (DAFFM). This method is built upon the Transformer attention mechanism and introduces a dual-path cross-attention structure that enhances the interaction and fusion between I/Q signal features and time-frequency features. By alternating the dominant mode mechanism, the model effectively captures complementary information between modes. Furthermore, DAFFM uses a self-attention module to enhance the intra-modal dependencies, which improves the overall feature extraction capability. Experimental results on the ORACLE dataset demonstrate the superior performance of the proposed method, achieving 85.16% recognition accuracy at SNR = 0 dB, which represents a 37.97% improvement over the state-of-art method.