Research map / Adversarial robustness and security / Certified and empirical robustness
Adversarial Robustness on In- and Out-Distribution Improves Explainability
Maximilian Augustin, Alexander Meinke, Matthias Hein
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add adversarial training on the out-distribution (ACET) to AT0.5 (RATIO-0.5) (ablation) | AT0.5 · CIFAR10 | robust accuracy l2 eps=0.5 (%): 70.42 -> 73.27 | Improved |
| add adversarial training on the out-distribution to AT0.5 (RATIO-0.5) (ablation) | AT0.5 · CIFAR10 | worst-case OOD AUC l2 eps=1.0 (%): 57.4 -> 83.6 | Improved |
| add adversarial training on the out-distribution to AT0.25 (RATIO-0.25) (ablation) | AT0.25 · CIFAR10 | worst-case OOD AUC l2 eps=1.0 (%): 42.0 -> 84.3 | Improved |
| add adversarial training on the out-distribution to AT0.25 (RATIO-0.25) (ablation) | AT0.25 · CIFAR10 | clean accuracy (%): 94.00 -> 93.53 | No clear effect |
| add adversarial training on the in-distribution to ACET (RATIO-0.25) (ablation) | ACET · CIFAR10 | worst-case OOD AUC l2 eps=1.0 (%): 81.9 -> 84.3 | Improved |
| add adversarial training on the out-distribution to AT0.5 (RATIO-0.5) (ablation) | AT0.5 · SVHN | robust accuracy l2 eps=0.5 (%): 68.58 -> 68.85 | No clear effect |
| add adversarial training on the out-distribution to AT0.5 (RATIO-0.5) (ablation) | AT0.5 · CIFAR100 | robust accuracy l2 eps=0.5 (%): 43.23 -> 45.55 | Improved |
| add adversarial training on the out-distribution to AT3.5 (RATIO-3.5) (ablation) | AT3.5 · Restricted ImageNet | worst-case OOD AUC l2 eps=7.0 (%): 37.5 -> 90.9 | Improved |
| use smaller in-distribution radius 0.25 instead of 0.5 for RATIO (ablation) | RATIO · CIFAR10 | clean accuracy (%): 91.08 -> 93.53 | Improved |
| use smaller in-distribution radius 0.25 instead of 0.5 for RATIO (ablation) | RATIO · CIFAR10 | robust accuracy l2 eps=0.5 (%): 73.27 -> 70.49 | Hurt |
| add out-distribution training and adversarial training (RATIO) instead of TRADES training | TRADES (own implementation) · CIFAR10 | robust accuracy l2 eps=0.5 (%): 71.02 -> 73.27 | Improved |
| train with outlier exposure only (OE) instead of plain training (ablation) | Plain · CIFAR10 | worst-case OOD AUC l2 eps=1.0 (%): 1.6 -> 8.7 | No clear effect |
Closest papers by meaning
- Robust Models are less Over-Confident (NeurIPS 2022 · similarity 0.88)
- On the Sensitivity of Adversarial Robustness to Input Data Distributions (ICLR 2019 · similarity 0.88)
- Fantastic Robustness Measures: The Secrets of Robust Generalization (NeurIPS 2023 · similarity 0.87)
- OODRobustBench: a Benchmark and Large-Scale Analysis of Adversarial Robustness under Distribution Shift (ICML 2024 · similarity 0.87)
- Adversarial Training Can Provably Improve Robustness: Theoretical Analysis of Feature Learning Process Under Structured Data (ICLR 2025 · similarity 0.87)
- Improving Accuracy-robustness Trade-off via Pixel Reweighted Adversarial Training (ICML 2024 · similarity 0.87)
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.
