CAAE: A Novel Wireless Spectrum Anomaly Detection Method with Multiple Scoring Criterion

Abstract

To sense and understand how to use the wireless spectrum, people have proposed various anomaly spectrum detection methods. We judge it as anomaly behavior if the received signal is unauthorized or the radiation of an expected signal is changed. We propose CAAE, a novel wireless spectrum anomaly detection method, to detect the two kinds of anomaly behaviors. CAAE is a complex adversarial autoencoder that can realize feature extraction and image reconstruction of input data through convolution and deconvolution networks. We train CAAE in a semi-supervised learning fashion and various values in the calculation process would change if the anomaly spectrum is input after the model training is completed. Therefore, we propose the multiple scoring criterion to help improve the detection accuracy of our model. The time-frequency waterfall graphs are input and we do two sets of experiments to prove the validity of our model. The experimental results show that the comprehensive detection capability of CAAE model is superior to the comparison algorithms for our dataset.