Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Distribution Prototype Diffusion Learning for Open-set Supervised Anomaly Detection

Fuyun Wang, Tong Zhang, Yuanzhi Wang, Yide Qiu, Xin Liu, Xu Guo, Zhen Cui

TL;DR

This work addresses open-set supervised anomaly detection (OSAD) where anomalies are scarce and normal data are abundant, making robust generalization to unseen anomalies essential.It introduces Distribution Prototype Diffusion Learning (DPDL), which jointly learns multiple Gaussian prototypes and a Schrödinger bridge to map normal features toward a prototype space while diffusing away from anomalies, effectively tightening normal boundaries.To enhance cross-domain generalization, Dispersion Feature Learning (DFL) projects features into a hyperspherical space with a von Mises–Fisher–style dispersion objective, increasing inter-sample separability and improving out-of-distribution detection.Empirical results on nine public datasets show state-of-the-art AUC under both general and hard settings, demonstrating strong generalization to unseen anomalies and robustness in few-shot anomaly scenarios, with the framework achieving notable improvements on several challenging industrial and medical datasets.

Abstract

In Open-set Supervised Anomaly Detection (OSAD), the existing methods typically generate pseudo anomalies to compensate for the scarcity of observed anomaly samples, while overlooking critical priors of normal samples, leading to less effective discriminative boundaries. To address this issue, we propose a Distribution Prototype Diffusion Learning (DPDL) method aimed at enclosing normal samples within a compact and discriminative distribution space. Specifically, we construct multiple learnable Gaussian prototypes to create a latent representation space for abundant and diverse normal samples and learn a Schrödinger bridge to facilitate a diffusive transition toward these prototypes for normal samples while steering anomaly samples away. Moreover, to enhance inter-sample separation, we design a dispersion feature learning way in hyperspherical space, which benefits the identification of out-of-distribution anomalies. Experimental results demonstrate the effectiveness and superiority of our proposed DPDL, achieving state-of-the-art performance on 9 public datasets.

Distribution Prototype Diffusion Learning for Open-set Supervised Anomaly Detection

TL;DR

This work addresses open-set supervised anomaly detection (OSAD) where anomalies are scarce and normal data are abundant, making robust generalization to unseen anomalies essential.It introduces Distribution Prototype Diffusion Learning (DPDL), which jointly learns multiple Gaussian prototypes and a Schrödinger bridge to map normal features toward a prototype space while diffusing away from anomalies, effectively tightening normal boundaries.To enhance cross-domain generalization, Dispersion Feature Learning (DFL) projects features into a hyperspherical space with a von Mises–Fisher–style dispersion objective, increasing inter-sample separability and improving out-of-distribution detection.Empirical results on nine public datasets show state-of-the-art AUC under both general and hard settings, demonstrating strong generalization to unseen anomalies and robustness in few-shot anomaly scenarios, with the framework achieving notable improvements on several challenging industrial and medical datasets.

Abstract

In Open-set Supervised Anomaly Detection (OSAD), the existing methods typically generate pseudo anomalies to compensate for the scarcity of observed anomaly samples, while overlooking critical priors of normal samples, leading to less effective discriminative boundaries. To address this issue, we propose a Distribution Prototype Diffusion Learning (DPDL) method aimed at enclosing normal samples within a compact and discriminative distribution space. Specifically, we construct multiple learnable Gaussian prototypes to create a latent representation space for abundant and diverse normal samples and learn a Schrödinger bridge to facilitate a diffusive transition toward these prototypes for normal samples while steering anomaly samples away. Moreover, to enhance inter-sample separation, we design a dispersion feature learning way in hyperspherical space, which benefits the identification of out-of-distribution anomalies. Experimental results demonstrate the effectiveness and superiority of our proposed DPDL, achieving state-of-the-art performance on 9 public datasets.

Paper Structure

This paper contains 22 sections, 22 equations, 3 figures, 6 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Preliminaries
  4. Method
  5. Problem Formulation
  6. Distribution Prototype Learning
  7. Dispersion Feature Learning
  8. Anomaly Score Prediction
  9. Experiment
  10. Dataset and Evaluation Metric
  11. Baselines
  12. Implementation Details
  13. Results under General Setting
  14. Results under the Hard Setting
  15. Ablation Study
  16. ...and 7 more sections

Figures (3)

  • Figure 1: Our proposed DPDL framework. It comprises three distinct modules: Distribution Prototype Learning (DPL, Sec. \ref{['sec:DPL']}), Dispersion Feature Learning (DFL, Sec. \ref{['sec:DFL']}), and anomaly score prediction (Sec. \ref{['sec:ASP']}). DPL transforms the distribution of normal samples to a space of learnable multiple Gaussian prototypes through building schrödinger bridge, meantime pushing anomaly distribution away from these prototypes. DFL operates in a hyperspherical space, enlarging the distances of intermediate features of all samples in a hyperspherical space to strengthen feature generalization for detecting anomalies. The score prediction module leverages a multi-instance-learning method to compute anomaly scores.
  • Figure 2: Ablation study for SB and DFL under the general settings and hard settings.
  • Figure 3: Parameter sensitivity analysis for $C$, $\epsilon$, $\kappa$ and $\lambda$.