LCR-GAN: Learning Crucial Representation for Anomaly Detection

Abstract

Anomaly detection is pivotal and challenging in artificial intelligence, which aims to determine whether a query sample comes from the same class, given a set of normal samples from a particular class. There are a plethora of anomaly detection methods based on generative models; however, these methods aim to make the reconstruction error of the training samples smaller or extract more information from the training samples. We believe that it is more important for anomaly detection to extract crucial representation from normal samples rather than more information, so we propose a semi-supervised method named LCR-GAN. We conducted extensive experiments on four image datasets and 15 tabular datasets to demonstrate the effectiveness of the proposed method. Meanwhile, we also carried out an anti-noise study to demonstrate the robustness of the proposed method.