Research map / Image generation / Autoregressive and latent generation
RODEO: Robust Outlier Detection via Exposing Adaptive Out-of-Distribution Samples
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 RODEO, which generates near-distribution, diverse outlier exposure data by CLIP-guided diffusion from inlier images, then adversarially trains an outlier detector on it to improve robust outlier detection.
Problem
Outlier detection performance collapses under adversarial attacks, largely because outlier exposure during adversarial training uses fixed, distant auxiliary outliers that do not teach robust features.
Approach
Generate outlier exposure data by starting a pretrained diffusion model from noised inlier images (partial denoising from a random step t0 in [0.3T, 0.6T]) and guiding it with CLIP toward near-outlier text labels (semantically similar labels plus negative-attribute prompts), with threshold-based filtering; then adversarially train a classifier with an extra outlier class.
Key results
- In novelty detection under PGD-1000, RODEO averages 63.5% AUROC versus 31.7% for PrincipaLS and 3.7% for Transformaly, with 83.7% clean AUROC.
- RODEO's adaptive OE reaches 85.1% clean / 65.6% PGD mean in the OE-technique comparison, above Dream-OOD (62.4 / 22.7) and GOE (64.2 / 29.4).
- The paper reports surpassing existing methods by up to 50% in adversarial AUROC in novelty detection.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use Adaptive OE instead of ImageNet OE (ablation) | RODEO · Novelty Detection, mean across datasets | AUROC (%) under PGD-1000: 36.5 -> 65.6 | Helped |
| use Adaptive OE instead of Gaussian noise OE (ablation) | RODEO · Novelty Detection, mean across datasets | AUROC (%) under PGD-1000: 7.1 -> 65.6 | Helped |
| use Adaptive OE instead of ImageNet Mixup (ablation) | RODEO · Novelty Detection, mean across datasets | AUROC (%) under PGD-1000: 25.3 -> 65.6 | Helped |
| use Adaptive OE instead of FITYMI (ablation) | RODEO · Novelty Detection, mean across datasets | AUROC (%) under PGD-1000: 20.6 -> 65.6 | Helped |
| use Adaptive OE instead of GOE (ablation) | RODEO · Novelty Detection, mean across datasets | AUROC (%) under PGD-1000: 29.4 -> 65.6 | Helped |
| use Adaptive OE instead of Dream-OOD (ablation) | RODEO · Novelty Detection, mean across datasets | AUROC (%) under PGD-1000: 22.7 -> 65.6 | Helped |
| use adaptive synthesized outliers rather than fixed OE (ablation) | RODEO · Novelty Detection, mean across datasets | FDC: 1.971 -> 3.325 | Helped |
| replace Tiny ImageNet OE with SVHN, MNIST, or Gaussian noise (ablation) | ALOE and ATD on CIFAR10 vs CIFAR100 · CIFAR10 vs CIFAR100 | Detection performance under adversarial attacks | Hurt |
Surprising findings
- Dream-OOD, trained on 5 billion samples, underperforms RODEO, which uses a 67-million-sample CLIP and diffusion backbone, attributed to embedding-space generation and LAION bias.
- Replacing Tiny ImageNet OE with SVHN, MNIST, or Gaussian noise noticeably degraded ALOE and ATD under adversarial attack on CIFAR10 vs CIFAR100.
Limitations
- Clean performance still lags behind state-of-the-art methods, reflecting the clean-adversarial trade-off.
- Requires access to inlier class label names to extract near-outlier text labels.
- Results on some datasets (e.g., MVTecAD) show low adversarial AUROC (14.9%).
Experimental setup
Datasets: CIFAR10 CIFAR100 MNIST FashionMNIST SVHN MVTecAD Head-CT BrainMRI Tumor Detection Covid19 ImageNet-30 ImageNet
Benchmarks: CIFAR10 CIFAR100 MNIST FashionMNIST SVHN MVTecAD Head-CT BrainMRI Tumor Detection Covid19 ImageNet-30
Metrics: AUROC FID Density Coverage FDC
Baselines: ViT-MSP ATOM ALOE ATD EXOE PLP Dream-OOD GOE FITYMI PatchCore
Builds on: Outlier Exposure adversarial training CLIP FITYMI Dream-OOD
Closest papers by meaning
- ALTBI: Constructing Improved Outlier Detection Models via Optimization of Inlier-Memorization Effect (AAAI 2025 · similarity 0.88)
- ODIM: Outlier Detection via Likelihood of Under-Fitted Generative Models (ICML 2024 · similarity 0.88)
- DOS: Diverse Outlier Sampling for Out-of-Distribution Detection (ICLR 2024 · similarity 0.87)
- Out-of-Distribution Detection with Prototypical Outlier Proxy (AAAI 2025 · similarity 0.87)
- Quantile-Based Maximum Likelihood Training for Outlier Detection (AAAI 2024 · similarity 0.87)
- Contamination-Resilient Anomaly Detection via Adversarial Learning on Partially-Observed Normal and Anomalous Data (ICML 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.