Memory-Augmented Generative Adversarial Networks for Anomaly Detection

We propose a memory-augmented deep learning model for semisupervised anomaly detection (AD). While many traditional AD methods focus on modeling the distribution of normal data, additional constraints in the modeling process are needed to distinguish between normal and abnormal data. The proposed model, named memory augmented generative adversarial networks (MEMGAN), is coupled with external memory units through attentional operations. One property of MEMGAN in the latent space is such that encoded normal data are expected to reside in the convex hull of the memory units, while the abnormal ones are separated outside. This property makes the AD process of MEMGAN more robust and reliable. Experiments on AD datasets adapted from MVTec, MNIST, CIFAR10, and Arrhythmia demonstrate that MEMGAN notably improves over previous AD models. We also find that the decoded memory units in MEMGAN are more diverse and interpretable than those in previous memory-augmented models.