DL-Based Anomaly Detection at the Physical-Layer of Cognitive Radio by Deep Support Vector Data Description

Aiming at achieving anomaly detection at the physical-layer of cognitive radio, a Deep Support Vector Data Description (Deep SVDD) based anomaly detection scheme is proposed in this paper. Specifically, two novel techniques with different sensitivities related to Signal-to-Noise Ratio (SNR)-walls are proposed for two typical application scenarios. On one hand, for studying the optimal boundary of the desired signal in the feature space, an Improved Deep SVDD (IDSVDD) is conceived for extracting low-dimensional features of the samples represented in the time-frequency domain. On the other hand, a joint algorithm of Deep SVDD and Modulation Classification (JDSMC) is developed, which is capable of extracting low-dimensional features utilizing deep Convolutional Neural Networks (CNN). Extensive simulation and experimental results demonstrate that: 1) the detection performance and the real-time performance of the proposed techniques outperform conventional control algorithms. 2) The proposed techniques can effectively detect abnormal signals in complex electromagnetic environments and effectively improve the self-aware ability of Cognitive Radio Network (CRN). 3) Over-the-Air (OTA) measurements using Software Defined Radio (SDR) in the lab validate the proposed techniques’ anomaly detection performance.