Finding a ‘New’ Needle in the Haystack: Unseen Radio Detection in Large Populations Using Deep Learning

Abstract

Radio frequency fingerprinting enhances security and privacy of wireless networks and communications by learning and extracting unique characteristics embedded in transmitted signals. Deep learning-based approaches learn radio fingerprints without hand-engineering features. One persisting drawback in deep learning methods is they identify only devices that are previously observed in a training set: if a radio signal from a new, unseen, device is passed through the classifier, the source device will be classified as one of the known devices. We propose a novel approach that facilitates new class detection without retraining a neural network, and perform extensive analysis of the proposed model both in terms of model parameters and real-world datasets. We accomplish this by first breaking down a longer transmission burst into smaller slices, and assessing classifier confidence on a new transmission based on per slice statistics: our approach detects a new device with 76% accuracy, while reducing the classification accuracy of 500 previously seen devices by no more than 10%.