RadioNet: Robust Deep-Learning Based Radio Fingerprinting

Radio fingerprinting identifies wireless devices by leveraging hardware imperfections embedded in radio frequency (RF) signals. While neural networks have been applied to radio fingerprinting to improve accuracy, existing studies are not robust due to two major reasons. First, there is a lack of informative parameter selections in pre-processing over RF signals. Second, deep-learning-based radio fingerprinting derives poor performance against temporal variations in the cross-day scenario. In this paper, we enhance the robustness of deep-learning-based radio fingerprinting from three aspects, including parameter selection in pre-processing, learning methods, and evaluation metrics. First, we conduct extensive experiments to demonstrate that careless selections of parameters in pre-processing can lead to over-optimistic conclusions regarding the performance of radio fingerprinting. Second, we leverage adversarial domain adaptation to improve the performance of radio fingerprinting in the cross-day scenario. Our results show that adversarial domain adaptation can improve the performance of radio fingerprinting in the cross-day scenario without the need of recollecting large-scale RF signals across days. Third, we introduce device rank as an additional metric to measure the performance of radio fingerprinting compared to using accuracy alone. Our results show that pursuing extremely high accuracy is not always necessary in radio fingerprinting. An accuracy that is reasonably greater than random guess could lead to successful authentication within a second when we measure with device rank.