Adversarial Learning-Based Spectrum Sensing in Cognitive Radio

In spectrum sensing, classical signal processing based sensing methods create a test statistic based on empirically statistical modeling. Recently, machine learning (ML) based methods use a neural network (NN) to learn a test statistic in a data-driven manner, but they can not well adapt to a new spectrum environment featured by a test signal-to-noise ratio (SNR) set with new SNR value(s). To address this issue, we propose a new adversarial learning based spectrum sensing method to improve the model adaptability. The key of our method is to design three coupled NNs, which can extract the universal less SNR-dependent features in the training SNR set, and use these features to infer the spectrum status in a new test SNR set. Simulation results show that the proposed method can achieve a significant performance improvement compared to the existing ML based methods and classical signal processing methods in terms of the spectrum sensing error rate.