Anomaly Detection using Generative Adversarial Networks Reviewing methodological progress and challenges

The applications of Generative Adversarial Networks (GANs) are just as diverse as their architectures, problem settings as well as challenges. A key area of research on GANs is anomaly detection where they are most often utilized when only the data of one class is readily available. In this work, we organize, summarize and compare key concepts and challenges of anomaly detection based on GANs. Common problems which have to be investigated to progress the applicability of GANs are identified and discussed. This includes stability and time requirements during training as well as inference, the restriction of the latent space to produce solely data from the normal class distribution, contaminated training data as well as the composition of the resulting anomaly detection score. We discuss the problems using existing work as well as possible (partial) solutions, including related work from similar areas of research such as related generative models or novelty detection. Our findings are also relevant for a variety of closely related generative modeling approaches, such as autoencoders, and are of interest for areas of research tangent to anomaly detection such as image inpainting or image translation.