Deep Learning Based RF Fingerprint Identification using Cyclic Shift Characteristic

Radio frequency fingerprint (RFF) is a promising technique to address Internet of Things (IoT) terminal authentications. We propose a novel RFF extraction method for signals with cyclic shift characteristics (CSC). During the preprocessing stage, a preamble spectrum correlation and signal tracking (PSCST) algorithm is proposed to compensate for carrier frequency offset. The robust RFF is extracted from the method based on cross power spectral density (CPSD), which utilizes all symbols within a frame. A convolutional neural network (CNN) is designed to recognize different devices using the CSC features. Compared to existing CNN-based RFF identification methods, the proposed CSC-CNN possesses the merits of data independency, channel interference resistance, and high communication efficiency, particularly in long-distance and low signal-to-noise ratio (SNR) conditions. The accuracy rate of CSC-CNN algorithm is higher than 98.42% in indoor/outdoor LOS scenario and 98.35% in NLOS scenario with a higher SNR. In a dynamic scenario where the terminal moves slowly in a circular path around the receiver, the highest accuracy attained through cross-time validation is 98.16%, which outperforms existing RFF identification methods when classifying 60 target LoRa devices.