Unsupervised Deep One-Class Classification with Adaptive Threshold based on Training Dynamics

TL;DR

This paper introduces an unsupervised deep one-class classification method that uses adaptive thresholds based on training dynamics to improve anomaly detection, especially when training data contains mixed normal and abnormal samples.

Contribution

It proposes a novel pseudo-labeling approach with adaptive thresholds to enhance deep one-class classification in practical scenarios with mixed data.

Findings

Significant performance improvements on 10 anomaly detection benchmarks.

Effective pseudo-labeling method enhances normality learning.

Robustness to mixed normal and abnormal samples in training data.

Abstract

One-class classification has been a prevailing method in building deep anomaly detection models under the assumption that a dataset consisting of normal samples is available. In practice, however, abnormal samples are often mixed in a training dataset, and they detrimentally affect the training of deep models, which limits their applicability. For robust normality learning of deep practical models, we propose an unsupervised deep one-class classification that learns normality from pseudo-labeled normal samples, i.e., outlier detection in single cluster scenarios. To this end, we propose a pseudo-labeling method by an adaptive threshold selected by ranking-based training dynamics. The experiments on 10 anomaly detection benchmarks show that our method effectively improves performance on anomaly detection by sizable margins.