A transformer-based method for radio-frequency fingerprinting of IoT devices

Radio Frequency Fingerprint Identification (RFFI) is a technique used to classify wireless devices by examining the unique signal distortions that arise from hardware imperfections inherent to each device. This method allows a wireless receiver to identify one or more transmitters accurately. Previous studies have presented RFFI results with wireless modulations such as LoRa, ZigBee, and Bluetooth Low Energy (BLE). This paper presents a method for RFFI in Internet of Things (IoT) devices using Gaussian Frequency Shift Keying (GFSK) modulation. The proposed RFFI method is based on an encoder module of a Vision Transformer (ViT); the results are compared with those obtained using a convolutional neural network (CNN). We present an analysis of the method’s accuracy in several propagation scenarios. We also analyze the effect of various parameters on the method’s accuracy, including training dataset size, training vector length, the portion of the transmitted packet to train on, and the number of epochs for training. Experimental findings indicate that the transformer-based classifier performs slightly better than a CNN (up to 5% better accuracy) in non-line-of-sight (NLOS) propagation conditions. More significantly, the transformer-based classifier requires only half the training epochs compared to the CNN.