Complex Neural Networks for Radio Frequency Fingerprinting

Abstract

We use deep learning to design a radio frequency (RF) fingerprint algorithm that takes complex-valued wireless signals as input, and outputs the identity of the device that transmitted the signal. We study how performance accuracy varies due to changes in input representation, choices of labels, and treatment of complex values. We report sensitivity to number of devices, training set size, signal-to-noise ratio, and environmental channel. Training data are real-time transmissions from thousands of devices.