“…Existing deep anomaly detection 1 methods [2,7,19,20,22,29,30] address these two challenges by using unsupervised deep learning to model the normal class in a two-step approach (i.e., the pipeline (a) in Figure 1): they first learn to represent data with new representations, e.g., intermediate representations in autoencoders [2,7,30], latent spaces in generative adversarial networks (GANs) [22,29], or distance metric spaces in [19,20]; and then they use the learned representations to define anomaly scores using reconstruction errors [2,7,22,29,30] or distance-based measures [19,20]. However, in most of these methods [2,7,22,29,30], the representation learning is separate from anomaly detection methods, so it may yield representations that are suboptimal or even irrelevant w.r.t. specific anomaly detection methods.…”