Research map / Label-efficient and robust learning / OOD and anomaly detection
An Evidence-Based Post-Hoc Adjustment Framework for Anomaly Detection Under Data Contamination
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes EPHAD, a post-hoc test-time adjustment that tilts unsupervised anomaly detector scores toward test-time evidence (CLIP, LOF, IForest) via exponential tilting, with an entropy-based adaptive temperature variant, improving detectors trained on contaminated data.
Problem
Unsupervised anomaly detectors assume clean training data, but real datasets contain undetected anomalies; existing remedies need training pipeline access, data, or known contamination ratios.
Approach
Treat the contaminated AD model's density/score as a prior and exponentially tilt it with a test-time evidence function T(x) (e.g., CLIP normality probability, LOF, IForest) weighted by temperature beta, yielding a revised inlier score that is closer in KL to the true inlier density; an entropy-minimisation rule (EPHAD-Ada) sets beta without labels.
Key results
- On the 10% contamination setting, EPHAD improved CFLOW on Fashion-MNIST from 72.87 to 92.97 AUROC and on CIFAR-10 from 65.47 to 97.38.
- COPOD with LOF evidence on tabular data improved on cover from 78.70 to 79.01 and on pendigits in some settings; LOF on cover improved from 22.44 to 44.20 with EPHAD.
- EPHAD-Ada, the entropy-based adaptive beta variant, often matches default beta=0.5 but does not always improve, and sometimes degrades performance (e.g., ULSAD on SVHN declined with EPHAD).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add EPHAD with CLIP evidence | CFLOW · FMNIST | AUROC: 72.87 { ($\pm$ 0.48)} -> 92.97 { ($\pm$ 0.19)} | Helped |
| Add EPHAD with CLIP evidence | CFLOW · RealIAD | AUROC: 76.42 { ($\pm$ 0.47)} -> 71.58 { ($\pm$ 0.17)} | Hurt |
| Use adaptive temperature in EPHAD with CLIP evidence | CFLOW · MVTec | AUROC: 87.58 { ($\pm$ 0.77)} -> 89.84 { ($\pm$ 0.3)} | Helped |
| Add EPHAD with CLIP evidence | ULSAD · SVHN | AUROC: 64.27 { ($\pm$ 0.22)} -> 58.17 { ($\pm$ 0.18)} | Hurt |
| Use adaptive temperature instead of the default EPHAD temperature | ULSAD + EPHAD with CLIP evidence · SVHN | AUROC: 58.17 { ($\pm$ 0.18)} -> 59.68 { ($\pm$ 0.16)} | Helped |
| Add EPHAD with LOF evidence | DeepSVDD · glass | AUROC: 64.52 { ($\pm$ 6.87)} -> 80.94 { ($\pm$ 3.31)} | Helped |
| Add EPHAD with LOF evidence | LOF · cover | AUROC: 22.44 { ($\pm$ 0.1)} -> 44.2 { ($\pm$ 0.07)} | Helped |
| Add EPHAD with LOF evidence | LOF · ionosphere | AUROC: 94.64 { ($\pm$ 0.52)} -> 89.74 { ($\pm$ 0.55)} | Hurt |
| Add EPHAD with LOF evidence | COPOD · pendigits | AUROC: 88.44 { ($\pm$ 0.2)} -> 88.38 { ($\pm$ 0.2)} | No clear effect |
| Vary contamination ratio from 5% to 15% (ablation) | EPHAD with CLIP evidence · Visual AD datasets | AUROC | No clear effect |
| Decrease temperature toward zero, increasing reliance on evidence (ablation) | EPHAD · Visual and tabular AD datasets | AUROC | Mixed |
| Use EPHAD-Ada instead of the best fixed-temperature EPHAD configuration (ablation) | EPHAD · Explored temperature range | AUROC | No clear effect |
Surprising findings
- Applying EPHAD with a weak evidence function can degrade strong detectors, e.g., ULSAD on SVHN, showing reliance on beta tuning.
- EPHAD-Ada does not consistently outperform fixed beta=0.5, and the authors hypothesise uncalibrated inlier probabilities are the cause.
Limitations
- Performance depends heavily on the temperature beta, and degradation occurs when evidence is much weaker than the AD model.
- The adaptive beta (EPHAD-Ada) is not consistently beneficial, likely due to uncalibrated inlier probabilities.
- Evidence functions used are generic; the framework's full potential needs domain-specific evidence design.
Experimental setup
Datasets: MVTecAD MPDD ViSA RealIAD CIFAR-10 Fashion-MNIST MNIST SVHN ADBench classical tabular datasets
Benchmarks: MVTecAD MPDD ViSA RealIAD MNIST Fashion-MNIST CIFAR-10 SVHN ADBench tabular (26 datasets)
Metrics: AUROC
Baselines: Refine Latent Outlier Exposure SoftPatch CLIP (WinCLIP) PatchCore PaDiM CFLOW FastFlow DRAEM Reverse Distillation ULSAD IForest
Builds on: DeepSVDD WinCLIP ADBench Local Outlier Factor
Closest papers by meaning
- ADPretrain: Advancing Industrial Anomaly Detection via Anomaly Representation Pretraining (NeurIPS 2025 · similarity 0.87)
- Contamination-Resilient Anomaly Detection via Adversarial Learning on Partially-Observed Normal and Anomalous Data (ICML 2024 · similarity 0.87)
- One-to-Normal: Anomaly Personalization for Few-shot Anomaly Detection (NeurIPS 2024 · similarity 0.86)
- UniVAD: A Training-free Unified Model for Few-shot Visual Anomaly Detection (CVPR 2025 · similarity 0.86)
- Long-Tailed Anomaly Detection with Learnable Class Names (CVPR 2024 · similarity 0.86)
- Generating and Reweighting Dense Contrastive Patterns for Unsupervised Anomaly Detection (AAAI 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.