Deep anomaly detection with self-supervised learning and adversarial training

Abstract

Deep anomaly detection, which utilizes neural networks to discover anomalies, is a vital research topic in pattern recognition. With the burgeoning of inference mechanism, inference-based methods show the promising performance. However, inference-based methods have two limitations: (1) they use an adversarial training way to learn data features. Such training way fails to learn task-specific features which can be conducive to capture the difference between normal and anomaly data. (2) The structure of detection network cannot capture the marginal distributions of normal data and corresponding features, which influences on the performance of anomaly detection. To overcome these limitations, this paper proposes a deep adversarial anomaly detection (DAAD) method. Specifically, an auxiliary task with self-supervised learning is first designed to learn task-specific features. Then a deep adversarial training (DAT) model is constructed to capture marginal distributions of normal data in different spaces. In addition, a majority voting strategy is applied to obtain reliable detection results. Experimental results on image and sequence datasets show that proposed method performs significantly better than many strong baselines.