Mixed CNN and Transformer for Specific Emitter Identification

Specific Emitter Identification (SEI) is a vital technique in radio sensing for identifying forged communication nodes and illegal base stations. In this study, we propose a mixed CNN-Transformer model that achieves state-of-the-art performance on datasets collected from real-world environments. First, we utilize large kernel expansion convolution for direct modeling of long time-series IQ data to avoid destroying the feature distribution of tiny RF damage by traditional hand-designed algorithms. Second, we design the Transformer module inspired by the Visual Transformer (ViT) architecture, which combines class labeling and absolute position encoding to facilitate global feature extraction. In a Bluetooth dataset collected from 33 mobile phones, our model achieves a Top-1 recognition rate of over 98.6% with four sampling rates. In experiments with a dataset of 16 WiFi devices, our data enhancement scheme is able to achieve excellent recognition rates in a small number of training sets, demonstrating its superior performance in SEI tasks in natural environments. All code is public: https://github.com/DuBianJun-007/Specific-Emitter-Identification.