A Specific Emitter Identification Method via Adaptive Labels-Directed Softmax to Increase Inter-Class Margin and Decrease Intra-Class Margin

Specific emitter identification (SEI) is considered an effective technical means to protect wireless networks from malicious intrusion, which identifies and classifies devices through radio frequency fingerprints (RFF) generated by hardware defects in transmitters. However, the differences between individual transmitters of the same type are extreme subtle, and the features available to distinguish them are extremely limited and fragile. This susceptibility to noise interference greatly affects the identification of transmitters. To address this issue, we propose a deep learning (DL) model that operates in the complex signals capable of extracting features from multiple domains and optimize the distribution of signals from different transmitters in the feature space to maximize the utilization of limited features. Specifically, in this paper, we introduce the U-fusion Transformer (UFformer) for efficient integration of multi-domain signals features, along with adaptive labels-directed softmax (ALD_Softmax) to optimize the backpropagation process, thereby reducing intra-class distances and increasing inter-class distances of radiation source signals to enhance classifier performance. Experimental results on real-world data demonstrate that our proposed model outperforms current SEI methods in terms of training cycles, recognition accuracy, and robustness.