DAS: Deep Autoencoder with Scoring Neural Network for Anomaly Detection

Abstract

Many anomaly detection methods are unsupervised e.g. they only utilize the non-anomalous data for model training. Data points that deviate from the majority of the pattern are deemed as anomalies. However, in many cases, anomaly labels are available which can help to guide the model learning for anomaly detection. We introduce an end-to-end anomaly score learning model composed of an autoencoder and a scoring neural network, which assigns an anomaly score to a given data point according to its level of abnormality. We jointly optimize the reconstruction loss and anomaly score function in an end-to-end manner. Experimental results on multiple datasets show that the proposed method appears to be superior over many state-of-the-art methods.