You May Need both Good-GAN and Bad-GAN for Anomaly Detection

Riqiang Gao, Zhoubing Xu, Guillaume Chabin, Awais Mansoor, Florin-Cristian Ghesu, Bogdan Georgescu, Bennett A. Landman, Sasa Grbic

Keywords: Anomaly Detection, GAN, Orthogonal Regularization, Bad-GAN

Abstract: Generative adversarial nets (GAN) have been successfully adapted for anomaly detection, where end-to-end anomaly scoring by so-called Bad-GAN has shown promising results. A Bad-GAN generates pseudo anomalies at the low-density area of inlier distribution, and thus the inlier/outlier distinction can be approximated. However, the generated pseudo anomalies from existing Bad-GAN approaches may (1) converge to certain patterns with limited diversity, and (2) differ from the real anomalies, making the anomaly detection hard to generalize. In this work, we propose a new model called Taichi-GAN to address the aforementioned issues of a conventional Bad-GAN. First, a new orthogonal loss is proposed to regularize the cosine distance of decentralized generated samples in a Bad-GAN. Second, we utilize few anomaly samples (when available) with a conventional GAN, i.e., so-called Good-GAN, to draw the generated pseudo anomalies closer to the real anomalies. Our Taichi-GAN incorporates Good-GAN and Bad-GAN in an adversarial manner; which generates pseudo anomalies that contributing to a more robust discriminator for anomaly scoring, and thus anomaly detection. Substantial improvements can be observed from our proposed model on multiple simulated and real-life anatomy detection tasks.

One-sentence Summary: Our proposed method incorporates a Good-GAN and a Bad-GAN in an adversarial manner for end-to-end anomaly detection.