Research map / Adversarial robustness and security / Certified and empirical robustness
Regularizing Hard Examples Improves Adversarial Robustness
Hyungyu Lee, Saehyung Lee, Ho Bae, Sungroh Yoon
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 DPLS (difficulty-based label smoothing) | PGD · CIFAR-10 A3 (eps=8) | robust accuracy: 51.80 -> 54.03 | Improved |
| add DPLS | TRADES · CIFAR-10 A3 (eps=8) | robust accuracy: 54.08 -> 55.19 | Improved |
| add DPLS | PGD · SVHN A3 (eps=8) | robust accuracy: 35.20 -> 42.01 | Improved |
| add DPLS | PGD · CIFAR-100 A3 (eps=8) | robust accuracy: 27.36 -> 29.33 | Improved |
| add uniform label smoothing (LS) with same budget | PGD · CIFAR-10 A3 (eps=8) | robust accuracy: 51.80 -> 51.99 | No clear effect |
| add hard-example pruning (top 10% difficulty) | PGD · CIFAR-10 A3 (eps=8) | robust accuracy: 51.80 -> 52.56 | Improved |
| prune top 10k hard examples from training (ablation) | PGD · CIFAR-10 | robust accuracy: 51.80 -> 53.36 | Improved |
| prune top 10k hard examples from training (ablation) | PGD · CIFAR-10 | standard accuracy: 87.19 -> 84.70 | Hurt |
| prune top 10k easy examples from training (ablation) | PGD · CIFAR-10 | robust accuracy: 51.80 -> 49.71 | Hurt |
| train only on 10k hard examples instead of 10k normal examples (ablation) | PGD · CIFAR-10 | robust accuracy: 39.56 -> 8.68 | Hurt |
| add DPLS and SWA (ablation) | PGD · CIFAR-10 A3 (eps=8) | robust accuracy: 53.98 -> 55.91 | Improved |
| use 0-1 loss for difficulty measure instead of EL2N (ablation) | DPLS on TRADES · CIFAR-10 A3 | robust accuracy: 54.32 -> 55.19 | Improved |
Closest papers by meaning
- Improving Adversarial Robustness Requires Revisiting Misclassified Examples (ICLR 2020 · similarity 0.91)
- Adversarial Training Can Provably Improve Robustness: Theoretical Analysis of Feature Learning Process Under Structured Data (ICLR 2025 · similarity 0.90)
- Soften to Defend: Towards Adversarial Robustness via Self-Guided Label Refinement (CVPR 2024 · similarity 0.90)
- A Closer Look at Curriculum Adversarial Training: From an Online Perspective (AAAI 2024 · similarity 0.88)
- Annealing Self-Distillation Rectification Improves Adversarial Training (ICLR 2024 · similarity 0.88)
- Wide Two-Layer Networks can Learn from Adversarial Perturbations (NeurIPS 2024 · similarity 0.88)
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.
