Anomaly Detection with Dual Adversarial Training
Shuo Liu, Liwen Xu · 2021 International Conference on Data Mining Workshops (ICDMW) · 2021
Anomaly detection is of paramount importance in data mining and artificial intelligence. Deep generative models have been widely used in anomaly detection as a dominant paradigm to model complex and high-dimensional data distribution. However, developing effective and robust anomaly detection systems for complex and high-dimensional data using generative models remains a challenge. In this paper, we propose a novel Dual Adversarial Training method for Anomaly Detection (DAT-AD), which uses the adversarial training idea in Generative Adversarial Network (GAN) and the Virtual Adversarial Training (VAT) idea to improve the effectiveness and robustness of anomaly detector, respectively. In addition, we have also carefully designed the network architecture and the loss function of our method to ensure that the trained network can be utilized to the greatest extent. We demonstrate the superiority of our method by conducting various experiments on tabular and image data.