Anomaly Detection for Radio Frequency Signals using Self-Supervised Deep Learning

Abstract

Detecting anomalous radio frequency (RF) data is a critical capability for spectrum monitoring, inference mitigation, and electronic warfare applications. In recent years, deep learning (DL) has shown promising results for the problem of RF anomaly detection; however, most of the work in the field assumes a priori knowledge about the data and that is not always practical. Labeling data can be time-consuming and expensive to collect, and knowledge of what anomalies the model might encounter are not usually available. Deep one-class classification (DOCC) using Deep Support Vector Data Description (DSVDD) has emerged as a promising method for learning compact representations of normal data and identifying deviations from expected signal behavior, especially in scenarios where anomalous data are unavailable or scarce. In this work DSVDD methods are applied to the problem of anomaly detection in digitally modulated RF signals and its effectiveness across multiple modulation schemes is demonstrated. Furthermore, a novel objective function based on Mahalanobis distance that models feature covariance in the learned latent space is introduced, addressing a key limitation of the original DSVDD method, which relies on Euclidean distance measures. Experimental results on simulated and over-the-air RF datasets show that the novel Mahalanobis method consistently outperforms the original DOCC approach. One limitation of DOCC methods is the assumption that the training data are all of the same class. The second half of this thesis relaxes this assumption and proposes a novel framework for fully self-supervised anomaly detection, training only on in-phase and quadrature (I/Q) streams of in-distribution signals, while never seeing examples of anomalies or having access to class labels in the training data. This framework consists of a two-step pipeline: first a contrastive learning step to cluster and segment the training data based on learned differentiating characteristics, and second a DSVDD step that learns to tightly compress the training data and exclude anomalies. This approach is shown to improve anomaly detection performance compared with using DSVDD without contrastive learning on multiple classes.