Boosting Semi-Supervised Anomaly Detection via Contrasting Synthetic Images

Sheng-Feng Yu, Wei-Chen Chiu

Abstract: In this paper we propose to tackle the problem of semi-supervised anomaly detection, which aims to learn the outlier detector from the training set composed of only inliers. Built upon the recent advances of introducing contrastive learning to achieve the state-of-the-art of anomaly detection, we propose a simple but effective extension to further boost the performance via integrating the contrastive learning and the generative model of inliers into a unified framework. On one hand, the contrastive learning amongst the real samples and synthetic ones produced by the generative model improves the representation learning; on the other hand, the generative model learning is also benefited from the contrastive learning. We conduct extensive experiments to demonstrate the efficacy of our proposed method to advance anomaly detection, its superiority against several baselines, and the contribution of our model designs.