Research map / Label-efficient and robust learning / OOD and anomaly detection
Beyond Clean Training Data: A Versatile and Model-Agnostic Framework for Out-of-Distribution Detection with Contaminated Training Data
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 a model-agnostic iterative weighting framework that lets existing OOD detectors train on contaminated (ID+OOD, unlabeled) data and simultaneously estimates the training set's OOD proportion.
Problem
OOD detection models are normally trained assuming a clean ID-only training set, but real training data is contaminated with unlabeled OOD samples, degrading detection; the OOD proportion of the training set is also unknown.
Approach
An iterative scheme where per-sample weights from the backbone's OOD scores are passed through a decaying transition function tau(t)=k/t+1 (keeping early-round weights high) and a sigmoid weight centered on an estimated ID ratio; the OOD ratio is re-estimated by thresholding moving-average scores. For negative-sample methods, weights are rescaled to [-1,1] so real OOD samples act as negatives.
Key results
- With 1% contamination on CIFAR-10 vs CelebA, the framework raised DB AUROC from 58.6 (backbone only) to 62.2 and VOS from 59.8 to 63.1, outperforming IAD.
- OOD proportion estimation error normalized by 0.5 was about 0.001-0.002 on ImageNet-1k and CIFAR-100 settings at 1% contamination, versus 0.980 for IAD's implicit 0.5 assumption.
- On CIFAR-10 vs CelebA at 5% contamination, the normalized estimation error remained at 0.004-0.016 across tested backbones.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply the proposed framework instead of IAD | DB, G-ODIN, and ReAct · CIFAR-100 → Tiny-ImageNet | AUROC: 55.9 -> 57.2; 58.7 -> 61.3; 70.1 -> 72.5 | Helped |
| Apply the proposed framework instead of IAD | DB, G-ODIN, and ReAct · CIFAR-100 → Mini-ImageNet | AUROC: 54.7 -> 55.6; 56.8 -> 60.4; 61.5 -> 64.4 | Helped |
| Apply the proposed framework instead of IAD | DB, G-ODIN, and ReAct · ImageNet-1k → iNaturalist | AUROC: 59.0 -> 60.7; 62.6 -> 64.8; 67.6 -> 70.3 | Helped |
| Apply the proposed framework instead of IAD | DB, G-ODIN, and ReAct · ImageNet-1k → OpenImage-O | AUROC: 53.8 -> 54.3; 55.6 -> 57.7; 60.8 -> 63.0 | Helped |
| Apply the proposed framework instead of IAD | VOS · CIFAR-100 → Tiny-ImageNet | AUROC: 58.5 -> 62.1 | Helped |
| Apply the proposed framework instead of IAD | VOS · CIFAR-100 → Mini-ImageNet | AUROC: 58.5 -> 59.6 | Helped |
| Apply the proposed framework instead of IAD | VOS · ImageNet-1k → iNaturalist | AUROC: 62.1 -> 63.2 | Helped |
| Apply the proposed framework instead of IAD | VOS · ImageNet-1k → OpenImage-O | AUROC: 54.3 -> 55.2 | Helped |
| Apply the proposed framework instead of IAD | DB and ReAct · CIFAR-10 → Celeb A, 1% contamination | AUROC: 60.1 -> 62.2; 69.1 -> 72.2 | Helped |
| Apply the proposed framework instead of IAD | DB and ReAct · CIFAR-10 → Celeb A, 2% contamination | AUROC: 58.4 -> 60.6; 66.7 -> 69.8 | Helped |
| Estimate the OOD percentage rather than fixing it at 0.5 (ablation) | Proposed framework with DB and VOS · CIFAR-10 → Celeb A, 1% contamination | AUROC: 61.0 -> 62.2; 61.9 -> 63.1 | Helped |
| Add the transition function τ(t) (ablation) | Proposed framework with DB and VOS · CIFAR-10 → Celeb A, 1% contamination | AUROC: 58.5 -> 62.2; 60.9 -> 63.1 | Helped |
Surprising findings
- Estimating the OOD proportion lets the method track the true contamination rate rather than assuming a fixed 50%, which IAD does implicitly.
Limitations
- Experiments use only 1-5% contamination in the main tables and two-dataset ID/OOD pairings.
- Termination threshold 0.05 and steepness hyperparameters chosen empirically.
- Not evaluated against methods designed specifically for contaminated data beyond IAD.
Experimental setup
Datasets: CIFAR-10 CelebA SVHN GTSRB CIFAR-100 Mini-ImageNet Tiny-ImageNet ImageNet-1k iNaturalist OpenImage-O
Benchmarks: CIFAR-10 vs CelebA CIFAR-100 vs Tiny-ImageNet CIFAR-100 vs Mini-ImageNet ImageNet-1k vs iNaturalist ImageNet-1k vs OpenImage-O
Metrics: AUROC normalized OOD proportion estimation error
Baselines: IAD BB only (backbone without framework)
Builds on: IAD (model-agnostic anomaly detection on contaminated data) curriculum learning VOS CSI
Closest papers by meaning
- Enhancing the Power of OOD Detection via Sample-Aware Model Selection (CVPR 2024 · similarity 0.91)
- Rethinking Out-of-Distribution Detection on Imbalanced Data Distribution (NeurIPS 2024 · similarity 0.89)
- Learning with Mixture of Prototypes for Out-of-Distribution Detection (ICLR 2024 · similarity 0.89)
- Image Background Serves as Good Proxy for Out-of-distribution Data (ICLR 2024 · similarity 0.89)
- Can We Ignore Labels in Out of Distribution Detection? (ICLR 2025 · similarity 0.89)
- MultiOOD: Scaling Out-of-Distribution Detection for Multiple Modalities (NeurIPS 2024 · similarity 0.89)
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.