An SNR-Aware Feature Reconstruction Method in Radio Frequency Fingerprint Identification

Abstract

The radio frequency fingerprint (RFF) has gained significant traction in the identification of wireless Internet of Things (IoT) devices. However, RFFs extracted from wireless signals are inherently susceptible to noise, particularly for narrowband signals. Furthermore, the noisy domain adaptation (NDA) problem presents a substantial challenge for RFF identification due to the variable noise interference across different noisy domains. To address this, the squared cross power spectral density (SCPSD) as new device RFFs is derived theoretically as a function of signal-to-noise ratio (SNR). Combined with the proposed high-precision SNR estimation algorithm, SCPSDs under low SNR can be reconstructed to the same feature distribution as those under high SNR. Because of the interpretability, ten samples under high SNR from each device under test (DUT) and a shallow convolutional neural network (CNN) are trained for experimental evaluation on the NDA problem. Tested on 60 off-the-shelf ZigBee DUTs, the improvement of identification accuracy is around 26% for SNR between 5 dB and 10 dB, and the overall improvement is more than 20% compared to the baseline. It outperforms the three other compared methods across all testing SNR and is highly practical.