Spectrum Anomaly Detection Under Multiple Sensors Using Metric-Adversarial Learning

Detecting spectrum anomaly effectively is vital for the spectrum security of critical regions. Existing deep learning-based spectrum anomaly detection (SAD) methods struggle to detect illegal users, which have unknown locations and frequencies. To overcome this limitation, we propose a multi-sensor SAD method based on deep metric learning (DML) and generative adversarial network (GAN). In this method, DML-GAN adopts a dual-metric approach to facilitate more robust anomaly detection. Besides, DML-GAN takes the matrix formed by the sensing results from multiple sensors as input. DML is adopted to extract semantic feature vectors and establish semantic center feature vectors for normal spectrum data. Subsequently, we calculate the Mahalanobis distance between these vectors as the first metric. GAN is employed to reconstruct the spectrum data, and the reconstruction error is chosen as the second metric. Simulation results show that the proposed method outperforms the existing methods.