ML-Based Detection Approaches for Covert Communication with Multi-D Signal Features

This work explores the efficacy of multidimensional (Multi-D) signal features in detecting covert communication, specifically considering time-domain features, frequency-time features, frequency offset, and SNR. To this end, we propose two novel machine learning (ML)-based detection schemes utilizing these signal features: a supervised learning approach employing k-nearest neighbors (KNN) and an unsupervised learning approach leveraging density-based spatial clustering of applications with noise (DBSCAN). To assess the performance of these schemes, we construct a covert communication testbed that enables the extraction of signal features in diverse electro-magnetic environments. Our experimental results show that both the KNN-based and DBSCAN-based schemes outperform traditional energy-based detection methods, and that our proposed schemes are more effective in identifying covert communication in complex electromagnetic environments.