Deep Learning for RF Fingerprinting: A Massive Experimental Study

Abstract

RF fingerprinting is a key security mechanism that allows device identification by learning unchanging, hardware-based characteristics of the transmitter. In this article, we demonstrate how machine learning techniques impact RF fingerprinting by analyzing a dataset of 400 GB of in-phase (I) and quadrature (Q) signal data transmitted by 10,000 radios. Our deep convolutional neural network architectures take raw and processed IQ samples as input to identify devices under a variety of practical conditions, including changing channels, noise levels, training data sizes, and computational overheads, among others. The contributions of the article are as follows: (i) This is the first work, to the best of our knowledge, reporting on RF fingerprinting and scalability issues on very large device populations, in the range of 50-10,000 devices. (ii) We investigate how to mitigate the effect of the wireless channel through a feature engineering step that we refer to as partial equalization. (iii) We provide a comprehensive performance evaluation of both a custom-designed and a modified ResNet architecture, with insights on which one may be preferred under specific environmental conditions.