Robust Signal Classification Using Unsupervised Learning

Abstract

Spectrum sensing is required for many cognitive radio applications, including spectral awareness, interoperability, and dynamic spectrum access. Any approach is highly dependent on the signal environment in which it operates. Noise, multipath channel distortion, sample-rate offsets, and center-frequency mis-estimations can make feature-based signal detection difficult. Additionally, as these values change in a dynamic environment, statically-trained classifiers cannot track the evolving class statistics. Unsupervised learning allows a cognitive radio system to evolve its classifier as the radio environment evolves, without the need for expert-annotated signal samples. While this allows improved classifier performance, previous work has demonstrated that unsupervised techniques developed without security in mind suffer from classifier manipulation through transmission of carefully crafted signals that shift decision boundaries. This paper develops countermeasures to the class manipulation attacks that mitigates their efficacy, and shows the robustness of unsupervised learning for signal classification in an evolving RF environment.