Multi-Normal Prototypes Learning for Weakly Supervised Anomaly Detection

TL;DR

This work tackles weakly supervised anomaly detection when normal data are multi-modal and unlabeled data may contain anomalies. It introduces a reconstruction-based framework that learns multiple normal prototypes via deep embedding clustering and contrastive learning, coupled with a likelihood-weighted reconstruction loss to resist anomaly contamination. A unified scoring module combines reconstruction error, latent features, and prototype similarity to produce anomaly scores, with end-to-end training aided by dynamic balancing of three losses. Experiments across 15 diverse datasets show state-of-the-art performance, strong generalization to unseen anomalies, and robustness to varying labeled-anomaly ratios, underscoring the practical value of multi-normal prototypes in real-world anomaly detection.

Abstract

Anomaly detection is a crucial task in various domains. Most of the existing methods assume the normal sample data clusters around a single central prototype while the real data may consist of multiple categories or subgroups. In addition, existing methods always assume all unlabeled samples are normal while some of them are inevitably being anomalies. To address these issues, we propose a novel anomaly detection framework that can efficiently work with limited labeled anomalies. Specifically, we assume the normal sample data may consist of multiple subgroups, and propose to learn multi-normal prototypes to represent them with deep embedding clustering and contrastive learning. Additionally, we propose a method to estimate the likelihood of each unlabeled sample being normal during model training, which can help to learn more efficient data encoder and normal prototypes for anomaly detection. Extensive experiments on various datasets demonstrate the superior performance of our method compared to state-of-the-art methods.

Figures (6)

• Figure 1: Motivations of the proposed multi-normal prototype learning (b). Compared with single-normal prototype learning (a), multi-normal prototype (b) can more effectively capture the diversity of normal data and estimate the probability of unlabeled samples being normal to resist anomaly contamination.
• Figure 2: Overview of the proposed model. The input $x_i$ from a labeled anomaly set $\mathcal{X_A}$ and an unlabeled set $\mathcal{X_U}$ are represented by blue and yellow shades, respectively. The output $score_i$ indicates the likelihood of a sample being an anomaly.
• Figure 3: Effects of varying the ratios of labeled anomalies.
• Figure 4: Effects of varying the number of normal prototypes
• Figure 5: Anomaly detection results on synthetic data. (a) The synthetic dataset comprises three Gaussian clusters representing normal samples and scattered anomalies. (b) Anomaly detection using multi-normal prototypes successfully captures the three distinct normal modes, assigning low anomaly scores to normal samples while effectively identifying anomalies. (c) Anomaly detection using a single normal prototype fails to adapt to the multi-modal nature of the normal data, resulting in misclassified normal samples and missed anomalies.