Research map / Adversarial robustness and security / Certified and empirical robustness
Eliminating Catastrophic Overfitting Via Abnormal Adversarial Examples Regularization
Runqi Lin, Chaojian Yu, Tongliang Liu
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 AAER regularization to RS-FGSM | RS-FGSM · CIFAR10 eps=16/255 | PGD-50-10 robust accuracy: 0.00 -> 23.87 | Improved |
| add AAER regularization to RS-FGSM | RS-FGSM · CIFAR10 eps=12/255 | PGD-50-10 robust accuracy: 0.00 -> 32.17 | Improved |
| add AAER regularization to N-FGSM | N-FGSM · CIFAR10 eps=32/255 | PGD-50-10 robust accuracy: 8.30 -> 12.97 | Improved |
| add AAER to N-FGSM at small noise | N-FGSM · CIFAR10 eps=8/255 | PGD-50-10 robust accuracy: 47.91 -> 48.31 | Improved |
| add AAER to RS-FGSM | RS-FGSM · CIFAR100 eps=8/255 | PGD-50-10 robust accuracy: 10.58 -> 25.31 | Improved |
| use only regularization part (i) number of AAEs (ablation) | RS-AAER · CIFAR10 eps=16/255 | robust accuracy: 23.87 -> 0.00 | Hurt |
| remove part (i) number term (ablation) | RS-AAER · CIFAR10 eps=16/255 | robust accuracy: 23.87 -> 12.56 | Hurt |
| remove part (ii) prediction-confidence term (ablation) | RS-AAER · CIFAR10 eps=16/255 | natural accuracy: 64.56 -> 56.65 | Hurt |
| only use part (iii) logits term (ablation) | RS-AAER · CIFAR10 eps=16/255 | robust accuracy: 23.87 -> 9.77 | Hurt |
| enlarge step size alpha (ablation) | RS-AAER · CIFAR10 eps=16/255 | robust vs natural accuracy | Mixed |
| increase lambda1 from 1.0 to 2.0 (ablation) | AAER · CIFAR10 eps=16/255 | robust accuracy | No clear effect |
| AAER training overhead vs FGSM | RS/N-FGSM · CIFAR10 training time | seconds per epoch: 11.0 -> 11.2 | No clear effect |
Closest papers by meaning
- Efficient local linearity regularization to overcome catastrophic overfitting (ICLR 2024 · similarity 0.88)
- Overfitting in adversarially robust deep learning (ICML 2020 · similarity 0.88)
- Consistency Regularization for Adversarial Robustness (AAAI 2022 · similarity 0.88)
- Taxonomy Driven Fast Adversarial Training (AAAI 2024 · similarity 0.87)
- Rethinking Invariance Regularization in Adversarial Training to Improve Robustness-Accuracy Trade-off (ICLR 2025 · similarity 0.86)
- Benign Overfitting in Adversarial Training of Neural Networks (ICML 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.
