Lightweight CNN-Based RF Fingerprint Recognition Method
Junhao Feng, Xiaogang Tang, Binquan Zhang, Yanjie Ren · 2023
Large network models, numerous parameters, and computational complexity are issues that the current deep learning RF fingerprint recognition approach must deal with since mobile terminals have a finite amount of storage for computations. Therefore, a lightweight convolutional neural network structure is suggested in this article (IQCNN). To decrease the platform's additional signal transformation processing, the IQCNN model uses the original IQ signal as input. It then constructs a convolutional layer in the signal IQ to extract IQ-related features. The IQ-related features are created by building convolutional layers in the signal's IQ direction, reducing the two-dimensional data into one dimension, extracting the signal's time-domain features using a maximum pooling layer, dimensionality-reducing with a fully-connected layer, and then achieving classification recognition. The IQCNN achieves an average recognition rate of 84.82% on a publicly available dataset with 16 USRP devices and signal-to-noise ratios of -10 to 20 dB, which is better than other deep learning methods and conventional SVM methods and has a lightweight feature for better engineering applications.