Unsupervised Anomaly Detection by Autoencoder with Feature Decomposition

Abstract

In unsupervised anomaly detection tasks, a crucial challenge is modeling the underlying structure of normal data without knowing the definition or ratio of anomalies. The introduction of robustness against anomalous data in autoencoder architecture is a significant research focus in order to address this challenge. In this paper, we propose a model implemented by an autoencoder with two decoders, called Feature Decomposition AutoEncoder (FDAE). It maps all data into a high-dimensional latent feature space. Many studies have proved RSR technology and RPCA technology to improve the performance of anomaly detection models. FDAE employs RSR and RPCA techniques in the latent space to decompose latent features into normal features and abnormal features, then decodes them separately using two decoders. Furthermore, we design an optimization strategy to enable FDAE to prioritize modeling the underlying structure of normal data from unlabeled data to reduce the interference caused by unknown anomalous data. We demonstrate the high performance of FDAE in unsupervised anomaly detection tasks through experiments on five public datasets. In addition, we study the variation of FDAE’s anomaly detection capability under different noise scenarios on the MNIST dataset.