Research map / Label-efficient and robust learning / OOD and anomaly detection
Fast Decision Boundary based Out-of-Distribution Detector
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 fDBD, a hyperparameter-free, auxiliary-model-free post-hoc OOD detector that scores samples by a regularized average of closed-form lower-bound feature distances to decision boundaries.
Problem
Post-hoc out-of-distribution detection for pre-trained classifiers is effective with feature-space methods, but these rely on auxiliary models built from training features, which adds inference cost.
Approach
Measure a penultimate-feature's L2 distance to each class decision boundary via a closed-form lower bound computed from the logits and pre-computable weight-difference norms, then regularize the averaged distance by the feature's deviation from the training-feature mean; ID samples sit further from boundaries than OOD samples at equal deviation levels.
Key results
- On CIFAR-10 cross-entropy ResNet-18, fDBD reduces average FPR95 of MSP by 29.69 points (a 48.85% relative error reduction) and Energy by 13.78 points (30.73% relative).
- fDBD adds no measurable latency over the softmax baseline (0.53 ms per image on Tesla T4), while the CW-attack distance estimate takes 992.2 ms per image per class.
- The closed-form distance estimate differs from CW-attack estimates by less than 1.5%, and fDBD outperforms baselines on the ImageNet OOD benchmark averaged over FPR95 and AUROC.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use fDBD instead of MSP | ResNet-18 trained with cross-entropy loss · CIFAR-10 OOD benchmark | average FPR95: 29.69% | Helped |
| Use fDBD instead of Energy | ResNet-18 trained with cross-entropy loss · CIFAR-10 OOD benchmark | average FPR95: 13.78% | Helped |
| Regularize average decision-boundary distance by distance to the training-feature mean (ablation) | fDBD · ImageNet OOD benchmark | AUROC | Helped |
| Use distance to the training-feature mean alone as the detection score (ablation) | fDBD · ImageNet OOD benchmark | AUROC | No clear effect |
| Use the unregularized average decision-boundary distance instead of regularized distance (ablation) | fDBD · ImageNet OOD benchmark | AUROC | Hurt |
| Increase the number of smallest decision-boundary distances averaged (ablation) | fDBD · CIFAR-10 and ImageNet OOD benchmarks | average FPR95 and AUROC | Helped |
| Combine fDBD with activation shaping using ReAct, ASH, or Scale (ablation) | fDBD on ImageNet ResNet-50 · ImageNet OOD benchmark | AUROC and FPR95 | Helped |
| Increase CIFAR-10-C corruption severity | fDBD · CIFAR-10 OOD benchmark | AUROC | Hurt |
| Use supervised contrastive loss rather than cross-entropy training | CIFAR-10 and ImageNet classifiers with fDBD · CIFAR-10 and ImageNet OOD benchmarks | OOD detection performance | Helped |
| Use fDBD rather than KNN | ResNet-50 · ImageNet OOD benchmark | average FPR95 and average AUROC | Helped |
| Use the closed-form distance estimate rather than iterative CW attack estimation | Feature distance measurement · CIFAR-10 classifier | distance estimation latency: 992.2ms | Helped |
Surprising findings
- Deviation from the training mean alone gives AUROC near 50, yet regularizing the decision-boundary distance by it substantially improves separation.
Limitations
- Relies on assumptions (Gaussian mixture with simplex ETF class means) for the theoretical justification of the regularization.
- Evaluated only on image classification with CIFAR-10 and ImageNet and a few architectures.
Experimental setup
Datasets: CIFAR-10 ImageNet SVHN iSUN Places365 Texture iNaturalist SUN CIFAR-10-C
Benchmarks: CIFAR-10 OOD benchmark (SVHN, iSUN, Places365, Texture) ImageNet OOD benchmark (Texture, Places365, iNaturalist, SUN) CIFAR-10-C
Metrics: FPR95 AUROC inference latency (ms)
Baselines: MSP ODIN Energy MDS KNN ViM CSI SSD+ KNN+ ReAct ASH Scale
Builds on: Mahalanobis distance OOD detection energy-based OOD detection KNN OOD detection ViM supervised contrastive learning
Closest papers by meaning
- Enhancing the Power of OOD Detection via Sample-Aware Model Selection (CVPR 2024 · similarity 0.90)
- Learning to Shape In-distribution Feature Space for Out-of-distribution Detection (NeurIPS 2024 · similarity 0.90)
- BOOD: Boundary-based Out-Of-Distribution Data Generation (ICML 2025 · similarity 0.89)
- Detecting Out-of-Distribution Through the Lens of Neural Collapse (CVPR 2025 · similarity 0.89)
- Splitting & Integrating: Out-of-Distribution Detection via Adversarial Gradient Attribution (ICML 2025 · similarity 0.89)
- Leveraging Perturbation Robustness to Enhance Out-of-Distribution Detection (CVPR 2025 · 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.