Anomaly Detection with Pipeline Structure using Joint Distribution

We presented a novel way for the problem of anomaly detection, where, given a set of examples, the goal is to judge if the type of a query example is out of the training dataset. The key contribution of our work is to improve the accuracy of GAN-based anomaly detection and extend the use cases of this type of method (such as audio). In order to accomplish this goal, firstly, we replaced generator of standard GAN with pipeline structure consisting of autoencoder. Secondly, using joint contribution to improve the quality of reconstruction. Thirdly, we used Self-Attention to capture the long-range dependence of the time series for detecting anomalies in our network. The effectiveness of the proposed method is measured across three available datasets including image as well as audio datasets, and the desired results are achieved.