Research map / Adversarial robustness and security / Certified and empirical robustness
Mitigating Catastrophic Overfitting in Fast Adversarial Training via Label Information Elimination
Chao Pan, Ke Tang, Qing Li, Xin Yao
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 |
|---|---|---|---|
| initialize with label-information perturbation instead of uniform init (ablation) | LIET · CIFAR-10 (eps 8/255) | AutoAttack robust accuracy: 5.60 -> 50.01 | Improved |
| initialize with label-information perturbation instead of Bernoulli init (ablation) | LIET · CIFAR-10 (eps 8/255) | AutoAttack robust accuracy: 48.29 -> 50.01 | Improved |
| initialize with label-information perturbation instead of uniform init at eps 16/255 (ablation) | LIET · CIFAR-10 (eps 16/255) | AutoAttack robust accuracy: 21.52 -> 25.22 | Improved |
| initialize with label-information perturbation instead of Bernoulli init at eps 16/255 (ablation) | LIET · CIFAR-10 (eps 16/255) | AutoAttack robust accuracy: 20.31 -> 25.22 | Improved |
| use non-uniform label smoothing instead of uniform (ablation) | LIET · CIFAR-10 (eps 8/255) | AutoAttack robust accuracy: 49.64 -> 50.01 | Improved |
| use non-uniform label smoothing instead of uniform (ablation) | LIET · CIFAR-10 (eps 16/255) | AutoAttack robust accuracy: 24.59 -> 25.22 | Improved |
| add adversarial initialization with plus vs minus sign (ablation) | FGSM-PGI · CIFAR-10 (eps 8/255) | PGD-50 robust accuracy: 55.69 vs 55.87 | No clear effect |
| add adversarial initialization with plus vs minus sign (ablation) | FGSM-UAP · CIFAR-10 (eps 8/255) | AutoAttack robust accuracy: 49.55 vs 49.47 | No clear effect |
| use LIET instead of FGSM-AT | LIET · CIFAR-10 (eps 8/255) | AutoAttack robust accuracy: 0.00 -> 50.01 | Improved |
| use LIET instead of FGSM-PGI | LIET · CIFAR-10 (eps 16/255) | AutoAttack robust accuracy: 0.11 -> 25.22 | Improved |
| use LIET instead of PGD-AT-WA at eps 8/255 | LIET · CIFAR-10 (eps 8/255) | AutoAttack robust accuracy: 50.26 -> 50.01 | No clear effect |
| use LIET instead of PGD-AT-WA on Tiny-ImageNet | LIET · Tiny-ImageNet (eps 8/255) | AutoAttack robust accuracy: 19.62 -> 19.79 | No clear effect |
Closest papers by meaning
- Preventing Catastrophic Overfitting in Fast Adversarial Training: A Bi-level Optimization Perspective (ECCV 2024 · similarity 0.91)
- Soften to Defend: Towards Adversarial Robustness via Self-Guided Label Refinement (CVPR 2024 · similarity 0.90)
- Adversarial Initialization with Universal Adversarial Perturbation: A New Approach to Fast Adversarial Training (AAAI 2024 · similarity 0.87)
- Taxonomy Driven Fast Adversarial Training (AAAI 2024 · similarity 0.86)
- Layer-Aware Analysis of Catastrophic Overfitting: Revealing the Pseudo-Robust Shortcut Dependency (ICML 2024 · similarity 0.86)
- Annealing Self-Distillation Rectification Improves Adversarial Training (ICLR 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.
