1. DA-ResNeXt50 method for radio frequency fingerprint identification based on time-frequency and bispectral feature fusion
- Author
-
CHEN Mengdi, ZHANG Wei, SHEN Lei, LEI Fuqiang, and ZHANG Jiafei
- Subjects
short-time Fourier transform ,bi-spectrum transform ,radio frequency fingerprint ,dense connection ,asymmetric convolution ,Telecommunication ,TK5101-6720 ,Technology - Abstract
To address the problems that a single feature in radio frequency fingerprint recognition could not fully represent the integrity of the signal and that the differences between features of different classes were small, which limited the recognition accuracy, a DA-ResNeXt50 (ResNeXt50 with dense connection and ACBlock) method for radio frequency fingerprint identification based on time-frequency and bi-spectral feature fusion was proposed. Firstly, short-time Fourier transform (STFT) and bi-spectrum transform were performed separately on the signals collected from different devices, the resulting images were bi-narized and then concatenated. By taking advantage of the advantages of both transformations in the time-frequency domain and high-order statistical characteristics respectively, the radio frequency fingerprint features of different devices can be extracted and characterized more comprehensively. Then, the DA-ResNeXt50 network model was proposed. Borrowing from the idea of dense connection, each layer of the four-layer residual unit was directly connected to all previous layers, promoting feature reuse and transmission, which enabled it to better capture subtle differences between classes. Finally, the asymmetric convolution block (ACBlock) was used to replace the 3×3 convolution in the last residual unit of the model. This effectively increased the receptive field of the network and enhanced the skeleton part of the convolutional kernel, thereby improving the performance of radio frequency fingerprint recognition. The experimental results show that compared with that of using a single feature extraction method, the proposed feature fusion approach significantly improves performance. Compared with various classical models, the improved model has higher recognition accuracy.
- Published
- 2024
- Full Text
- View/download PDF