[Paper Review] What is Wrong with One-Class Anomaly Detection?

Junekyu Park, Jeong-Hyeon Moon+2 more|arXiv (Cornell University)|Apr 20, 2021

Anomaly Detection Techniques and Applications25 references6 citations

TL;DR

This paper identifies a critical flaw in one-class anomaly detection: existing methods fail when normal data contains diverse semantic classes, leading to overly loose decision boundaries. To address this, the authors propose CLAD, a confidence-based self-labeling framework that leverages latent class information via unsupervised clustering and pseudo-labeling, significantly outperforming prior one-class methods in multi-class normal scenarios.

ABSTRACT

From a safety perspective, a machine learning method embedded in real-world applications is required to distinguish irregular situations. For this reason, there has been a growing interest in the anomaly detection (AD) task. Since we cannot observe abnormal samples for most of the cases, recent AD methods attempt to formulate it as a task of classifying whether the sample is normal or not. However, they potentially fail when the given normal samples are inherited from diverse semantic labels. To tackle this problem, we introduce a latent class-condition-based AD scenario. In addition, we propose a confidence-based self-labeling AD framework tailored to our proposed scenario. Since our method leverages the hidden class information, it successfully avoids generating the undesirable loose decision region that one-class methods suffer. Our proposed framework outperforms the recent one-class AD methods in the latent multi-class scenarios.

Motivation & Objective

• To identify the fundamental limitation of one-class anomaly detection when normal data contains diverse semantic classes.
• To propose a realistic latent class-condition-based anomaly detection scenario where normal data comprises multiple hidden semantic classes.
• To develop a self-supervised framework that leverages latent class information without requiring ground-truth labels.
• To bridge the gap between supervised and unsupervised anomaly detection by enabling semantic-aware decision boundaries in the absence of abnormal labels.
• To demonstrate that learning from latent class structure improves anomaly detection performance over standard one-class methods.

Proposed method

• The method uses a self-supervised feature encoder trained via contrastive learning (SimCLR) to extract meaningful latent representations.
• It applies the Deep Embedding Clustering (DEC) algorithm to cluster the latent features and assign pseudo-labels to normal samples.
• A classifier (ResNet-18) is trained in a supervised fashion on these pseudo-labeled normal samples to learn class-aware representations.
• At inference, the model uses confidence-based anomaly detection (ODIN) with temperature scaling and input perturbation to detect anomalies.
• The framework avoids the loose decision boundary problem by learning tight, class-wise boundaries through latent semantic supervision.
• The entire pipeline is trained end-to-end with contrastive pre-training, clustering-based self-labeling, and supervised classification.

Experimental results

Research questions

• RQ1 Why do standard one-class anomaly detection methods fail when normal data contains multiple semantic classes?
• RQ2 Can latent class information be effectively leveraged in anomaly detection without access to ground-truth labels?
• RQ3 Can a self-labeling framework based on clustering improve decision boundary tightness and detection performance in one-class AD?
• RQ4 How does the proposed method compare to state-of-the-art one-class anomaly detection methods in multi-class normal scenarios?
• RQ5 Does confidence-based inference with temperature scaling and input perturbation enhance anomaly detection in the proposed framework?

Key findings

• The proposed CLAD framework significantly outperforms recent one-class anomaly detection methods on latent multi-class scenarios, especially in challenging cases like MNIST-Mixed.
• The method achieves higher AUC scores than baseline one-class methods across all datasets, including MNIST, GTSRB, CIFAR-10, and Tiny-ImageNet.
• The use of self-labeled pseudo-classes enables the model to learn tighter, class-wise decision boundaries, reducing false negatives caused by loose coverage.
• The ablation study confirms that self-labeling via clustering is essential for performance gains, as removing it degrades results to baseline levels.
• Temperature scaling and input perturbation in inference further improve detection confidence and AUC, validating the effectiveness of the ODIN-based confidence calibration.
• The framework generalizes well across diverse datasets, demonstrating robustness in real-world scenarios with heterogeneous normal classes.