Deep Learning-Based Anomaly Detection using Hybrid Loss

Abstract

Recently, deep learning-based (DL) automatic modulation classification (AMC) has been extensively studied in cooperative and non-cooperative communication contexts such as cognitive radio and spectrum surveillance. One of the drawbacks of DL-based AMC is its susceptibility to anomalous or interfering signals. In this paper, we propose a DL-based anomaly detection for AMC, utilizing an autoencoder to process the in-phase and quadrature components of a received signal. In order to detect anomalies, we employ a hybrid loss, a combination of the autoencoder's reconstruction loss and the Mahalanobis distance of the latent space embedding of the training vector and each input instance. Through computer simulations, we show that the proposed model has superior detection performance with less computational complexity compared to the conventional DL-based model.