Anomaly Detection in High-Dimensional Data Based on Autoregressive Flow

Abstract

Anomaly detection of high-dimensional data is an important but yet challenging problem in research and application domains. Unsupervised techniques typically rely on the density distribution of the data to detect anomalies, where objects with low density are considered to be abnormal. The state-of-the-art methods solve this problem by first applying dimension reduction techniques to the data and then detecting anomalies in the low dimensional space. However, these methods suffer from inappropriate density estimation modeling and decoupled models with inconsistent objectives. In this work, we propose an effective Anomaly Detection model based on Autoregressive Flow (ADAF). The key idea is to unify the distribution mapping capability of flow-based models with the neural density estimation power of autoregressive models. We design an autoregressive flow-based model to infer the latent variables of input data by minimizing the combination of latent error and neural density. The neural density of input data can be estimated naturally by ADAF, along with the latent variable inference, rather than through an additional stitched density estimation network. Unlike stitching decoupled models, ADAF optimizes the same network parameters simultaneously by balancing latent error and neural density estimation in a unified training fashion to effectively separate the anomalies out. Experimental results on six public benchmark datasets show that, ADAF achieves better performance than state-of-the-art anomaly detection techniques by up to 20% improvement on the standard \(F_{1}\) score.