Cooperative Deep Unsupervised Anomaly Detection

Abstract

In last years deep learning approaches to anomaly detection are becoming very popular. In most of the first methods the paradigm is to train neural networks initially designed for compression (Auto Encoders) or data generation (GANs) and to detect anomalies as a collateral result. Recently new architectures have been introduced in which the expressive power of deep neural networks is associated with objective functions specifically designed for anomaly detection. One of these methods is $\textit{Deep-SVDD}$ which, although created for One-Class classification, has been successfully applied to the (semi-)supervised anomaly detection setting. Technically, $\textit{Deep-SVDD}$ technique forces the deep latent representation of the input data to be enclosed into an hypersphere and labels as anomalies data farthest from its center. In this work we introduce $\textit{Deep-UAD}$, a neural network approach for unsupervised anomaly detection where, iteratively, a network similar to that of $\textit{Deep-SVDD}$ is alternatively trained with an Auto Encoder and the two networks share some weights in order for each network to improve its training by exploiting the information coming from the other network. The experiments we conducted show that the performances obtained by the proposed method are better than the ones obtained both by deep learning methods and standard shallow algorithms.