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.

Automated summary

The paper investigates the issue of spectrum sensing, where classical methods and machine learning methods fail to adapt to new signal-to-noise ratio environments. The authors propose a new adversarial learning method that improves model adaptability by designing coupled neural networks. Simulation results demonstrate significant improvement in spectrum sensing error rate compared to existing methods.