Quantile Autoencoder for Anomaly Detection

Hogeon Seo, Seunghyoung Ryu, Jiyeon Yim, Junghoon Seo, Yonggyun Yu

Keywords: Anomaly Detection, Novelty Detection, Outlier Detection, Autoencoder, Quantile

TL;DR: The proposed QAE using both the reconstruction error and the aleatoric uncertainty term in anomaly scoring improves the AD performance.

Abstract: Anomaly detection (AD) is an essential task in a variety of industrial fields. AD based on deep neural networks has shown effective performance. Most methods for deep anomaly detection (DAD) use $\textit{the difference between the input and reconstructed data}$ or $\textit{the distance from the center of the cluster defined by normal cases}$ as a measure of abnormality. However, these metrics do not consider the diversity of normal cases. We propose a quantile autoencoder (QAE) as a novel DAD method to consider the data-oriented uncertainty. QAE obtains the anomaly score from both the reconstruction error and the channel-wise data uncertainty that is the range of the two quantiles of the reconstruction distribution. This anomaly scoring makes the score distributions of the normal and abnormal samples farther apart by narrowing the width of the distributions, which contributes to the improvement of AD performance. The performance of the proposed QAE was verified with various datasets, and the results show higher performance compared to the benchmark results.