Abstract

Anomaly detection has been a lasting yet active research area for decades. However, the existing methods are generally biased towards capturing the regularities of high-density normal instances with insufficient learning of peripheral instances. This may cause a failure in finding a representative description of the normal class, leading to high false positives. Thus, we introduce a novel anomaly detection model that utilizes a small number of labeled anomalies to guide the adversarial training. In particular, a weighted generative model is applied to generate peripheral normal instances as supplements to better learn the characteristics of the normal class, while reducing false positives. Additionally, with the help of generated peripheral instances and labeled anomalies, an anomaly score learner simultaneously learns (1) latent representations of instances and (2) anomaly scores, in an end-to-end manner. The experimental results show that our model outperforms the state-of-the-art anomaly detection methods on four publicly available datasets, achieving improvements of 6.15%-44.35% in AUPRC and 2.27%-22.3% in AUROC, on average. Furthermore, we applied the proposed model to a real merchant fraud detection application, which further demonstrates its effectiveness in a real-world setting.