AI-based RF-Fingerprinting Framework and Implementation using Software-Defined Radios

Abstract

Radio frequency (RF) fingerprinting is considered to be a promising security solution for wireless communications at the physical layer. RF fingerprinting is still in its infancy, and much research is needed to further improve the detection capabilities. To address this problem, in this paper, we propose utilizing software-defined radios (SDRs), which have proven to be extremely beneficial to the RF research community. We demonstrate the capability of RF fingerprinting by identifying the transmit radios that are in the pre-selected whitelist (authorized) and reject any other transmit radios not found in the whitelist. We have experimented with four different universal software-radio peripherals (USRPs) models with a total of fourteen USRPs for our RF fingerprinting solution. Deep learning models and transfer learning are used to train the RF fingerprinting models. Experimental results reveal that the ability of RF fingerprinting the USRPs drops as the hardware quality of USRPs improves. For low-end USRPs an accuracy of 99% is achieved; however, for high-end radios, the accuracy decreased to as low as 43%. This is due to the difficulty of finding anomalies with high-quality hardware, which is essential for successful RF fingerprinting.