ATDAD: One-class adversarial learning for tabular data anomaly detection

Abstract

The detection of tabular data anomalies is widely used in cybersecurity, medicine, and other fields. However, modelling tabular data is challenging. Unlike images and videos, tabular data have unknown structures and are of mixed type. Although most anomaly detection methods focus on images and videos, we propose a one-class adversarial learning method for tabular data anomaly detection (ATDAD). ATDAD changes the structure of generative adversarial networks (GANs) to better model high-dimensional complex tabular data. ATDAD adds dual encoders to stabilize the training and ensure cycle consistency of the samples and their reconstructions. ATDAD does not require any anomalous samples as prior knowledge in either the training or testing phases. To the best of our knowledge, ATDAD is the first study that focuses on the detection of tabular data anomalies using a GAN. Experiments show that ATDAD outperforms existing unsupervised anomaly detection methods for tabular data.