Data-Driven Wireless Anomaly Detection Using Spectral Features

Abstract

In this work, we present an anomaly detection algorithm for wireless spectrum data and evaluate its ability to accurately detect interfering transmissions. The algorithm considers three types of interfering signals: a co-channel transmission, an interfering continuous-wave transmission, or an adversarial replay attack. At the core of each anomaly detector is a feature space transformation, which learns a compressed representation of the data, followed by various metrics on this representation. Four feature space transformations and four detection metrics are investigated, each of which represents different characterizations of the distribution of the data, totaling 16 configurations of the anomaly detector. Our algorithm is evaluated using Wi-Fi spectrum data from real-world radio frequency (RF) captures in both mild and harsh channel conditions. We present an analysis of the performance of the various configurations under each of the interference types at varying signal-to-interference ratios. Lastly, we discuss the implications of this performance on the the distribution of the data and the recommended model to detect each type of anomaly.