Research map / Adversarial robustness and security / Certified and empirical robustness
Probabilistic Margins for Instance Reweighting in Adversarial Training
Qizhou Wang, Feng Liu, Bo Han, Tongliang Liu, Chen Gong, Gang Niu, Mingyuan Zhou, Masashi Sugiyama
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 |
|---|---|---|---|
| Replace LPS-based reweighting with PM-based reweighting | AT, MART, and TRADES · CIFAR-10 | AA accuracy: 0.70-7.93% | Improved |
| Replace LPS-based reweighting with PM-based reweighting | AT, MART, and TRADES · CIFAR-10 | PGD accuracy: 0.14-2.00% | Improved |
| Use adversarial PM rather than natural PM | MAIL · CIFAR-10 | AA accuracy: 44.25 \pm 0.45 vs 44.10 \pm 0.21; 50.60 \pm 0.22 vs 49.52 \pm 0.11 | Mixed |
| Use normalized probability-space margin instead of logit-space multi-class margin (ablation) | TRADES with MAIL · CIFAR-10 | AA accuracy: 49.20 \pm 0.09 -> 50.60 \pm 0.22 | Improved |
| Use sigmoid rather than hinge weight assignment (ablation) | TRADES with MAIL · CIFAR-10 | AA accuracy: 43.40 \pm 0.22 -> 50.60 \pm 0.22 | Improved |
| Use sigmoid rather than step weight assignment (ablation) | TRADES with MAIL · CIFAR-10 | AA accuracy: 50.30 \pm 0.20 -> 50.60 \pm 0.22 | Improved |
| Use MAIL-AT rather than standard AT | ResNet-18 · CIFAR-10 | PGD accuracy: 48.91 \pm 0.14 -> 55.25 \pm 0.23 | Improved |
| Use MAIL-TRADES rather than standard TRADES | ResNet-18 · CIFAR-10 | AA accuracy: 48.10 \pm 0.26 -> 50.60 \pm 0.22 | Improved |
| Use MAIL-AT rather than GAIRAT | ResNet-18 · CIFAR-10 | CW accuracy: 39.86 \pm 0.08 -> 48.88 \pm 0.11 | Improved |
| Choose MAIL-AT rather than MAIL-TRADES | MAIL · CIFAR-10, ResNet-18 | PGD accuracy: 53.68 \pm 0.14 -> 55.25 \pm 0.23 | Improved |
| Choose MAIL-TRADES rather than MAIL-AT | MAIL · CIFAR-10, ResNet-18 | AA accuracy: 44.22 \pm 0.21 -> 50.60 \pm 0.22 | Improved |
| Use MAIL-AT rather than GAIRAT | ResNet-18 · SVHN | AA accuracy: 37.00 \pm 0.30 -> 41.12 \pm 0.30 | Improved |
Closest papers by meaning
- Geometry-aware Instance-reweighted Adversarial Training (ICLR 2021 · similarity 0.88)
- Towards Million-Scale Adversarial Robustness Evaluation With Stronger Individual Attacks (CVPR 2025 · similarity 0.87)
- Doubly Robust Instance-Reweighted Adversarial Training (ICLR 2024 · similarity 0.87)
- Detecting Brittle Decisions for Free: Leveraging Margin Consistency in Deep Robust Classifiers (NeurIPS 2024 · similarity 0.85)
- Improving Accuracy-robustness Trade-off via Pixel Reweighted Adversarial Training (ICML 2024 · similarity 0.85)
- Treatment of Statistical Estimation Problems in Randomized Smoothing for Adversarial Robustness (NeurIPS 2024 · similarity 0.85)
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.
