Radio Frequency Fingerprinting Improved by Statistical Noise Reduction

Abstract

Radio frequency (RF) fingerprinting as an important non-password authentication technology has received extensive attention in recent years. It can be used as an underlying intrusion detection component for various wireless networks to enhance their wireless security further. In order to achieve more robust RF fingerprinting with deep learning, this paper develops a practical signal data augmentation solution consisting of two steps: 1) construction of sampling points 2) smooth filtering, which enables to enhance discriminability for individual radio examples. The underlying idea is that using a relatively long signal can implicitly perform noise reduction in a statistical sense since noise is often uncorrelated or partially correlated. As more sampling points accumulated, RF fingerprints could be more discernible. Experimental results have demonstrated that our proposed data augmentation can significantly improve recognition accuracy, especially in low signal-to-noise ratio (SNR) and small example conditions.