Research map / Label-efficient and robust learning / OOD and anomaly detection
Filter or Compensate: Towards Invariant Representation from Distribution Shift for Anomaly Detection
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 FiCo, which adds a distribution-specific compensation module and a distribution-invariant filter module to reverse distillation to learn distribution-invariant normality for anomaly detection under distribution shift.
Problem
Anomaly detection methods trained on in-distribution data degrade sharply under distribution shift (e.g., corruptions like brightness, contrast, blur, noise), and the teacher-student misalignment in reverse distillation causes detection failures.
Approach
Compensate distribution-specific information in the student network via residual DiSCo modules so teacher and student align on OOD inputs, and filter all abnormal (anomalous and distribution-specific) information with a DiIFi module that maps shallower student features toward the final-block compensated features; DiIFi is discarded at inference.
Key results
- On MVTec, FiCo achieves 97.6% average AUROC versus 97.0% for GNL (ID 98.8%, OOD scenarios averaging about 97.5%).
- On PACS, FiCo achieves 72.6% average AUROC, 2.94% above GNL's reported 69.7% reimplementation/70.9% reported figures.
- On CIFAR-10, FiCo achieves 71.1% average AUROC versus 70.4% for GNL.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FiCo instead of GNL | Anomaly detection · MVTec (all scenarios, average) | sample-level AUROC: 97.0 -> 97.6 | Helped |
| Use FiCo instead of GNL | Anomaly detection · PACS (all domains, average) | sample-level AUROC: 70.9 -> 72.6 | Helped |
| Use FiCo instead of GNL | Anomaly detection · CIFAR-10 (all scenarios, average) | sample-level AUROC: 70.4 -> 71.1 | Helped |
| Add DiSCo and L_Co to GNL (ablation) | GNL on PACS · PACS (all domains, average) | sample-level AUROC: 69.7 -> 70.1 | Helped |
| Add DiIFi and L_mse to DiSCo (ablation) | DiSCo on PACS · PACS (all domains, average) | sample-level AUROC: 70.1 -> 71.7 | Helped |
| Add L_nor to DiSCo + DiIFi (ablation) | DiSCo + DiIFi on PACS · PACS (all domains, average) | sample-level AUROC: 71.7 -> 72.6 | Helped |
| Vary alpha while holding other hyperparameters at their optimal values (ablation) | FiCo · MVTec | sample-level AUROC: 0.55% | No clear effect |
| Vary beta while holding other hyperparameters at their optimal values (ablation) | FiCo · MVTec | sample-level AUROC: 0.14% | No clear effect |
| Vary gamma while holding other hyperparameters at their optimal values (ablation) | FiCo · MVTec | sample-level AUROC: 0.61% | No clear effect |
Limitations
- Inference still relies on test-time augmentation (EFDM), adding cost
- Gains on some OOD scenarios (e.g., PACS sketch, CIFAR-10 gaussian noise) remain weak
- Requires augmentation-based consistency, and hyperparameters (alpha, beta, gamma) are tuned per dataset
Experimental setup
Datasets: MVTec PACS CIFAR-10
Benchmarks: MVTec PACS CIFAR-10
Metrics: sample-level AUROC
Baselines: Deep SVDD f-AnoGAN KD RD PatchCore RD++ SimpleNet PNI GNL RealNet
Builds on: Reverse Distillation GNL EFDM DyConv
Closest papers by meaning
- Unlocking the Potential of Reverse Distillation for Anomaly Detection (AAAI 2025 · similarity 0.84)
- ADPretrain: Advancing Industrial Anomaly Detection via Anomaly Representation Pretraining (NeurIPS 2025 · similarity 0.84)
- ResAD: A Simple Framework for Class Generalizable Anomaly Detection (NeurIPS 2024 · similarity 0.84)
- Distribution Prototype Diffusion Learning for Open-set Supervised Anomaly Detection (CVPR 2025 · similarity 0.84)
- When and How Does In-Distribution Label Help Out-of-Distribution Detection? (ICML 2024 · similarity 0.83)
- Beyond Normal References: Discriminative Few-Shot Anomaly Detection (NeurIPS 2026 · similarity 0.83)
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.