Research map / Adversarial robustness and security / Certified and empirical robustness
Catastrophic Overfitting: A Potential Blessing in Disguise
MN Zhao, Lihe Zhang, Yuqiu Kong, Baocai Yin
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 feature-activation-difference suppression loss at node M_B (alpha3 200) (ablation) | FGSM-MEP fast adversarial training · CIFAR-10 ResNet18, train 12/255 | clean accuracy / AA: stable at 74.08 clean, 19.72 AA | Improved |
| suppress feature differences at node A, C, D or E instead of B (ablation) | FGSM-MEP · CIFAR-10 ResNet18, train 12/255 | CO avoidance: CO persists at nodes A, C, D | No clear effect |
| use random uniform noise at inference on CO-affected model | CO-affected ResNet18 · CIFAR-10 train 16/255 | AutoAttack accuracy (%): PGD-50 58.48, AA 43.53 with clean 88.45 | Improved |
| add input noise at inference to a clean-trained model (ablation) | ResNet18 clean-trained · CIFAR-10 | PGD-50 accuracy (%): 0.00 -> 0.00 | No clear effect |
| add input noise at inference to a clean-trained model (ablation) | ResNet18 clean-trained · CIFAR-10 | clean accuracy (%): 94.40 -> 80.63 | Hurt |
| add input noise at inference to CO model | CO-affected ResNet18 · CIFAR-10 clean accuracy | clean accuracy (%): 89.81 -> 89.64 | No clear effect |
| use uniform noise instead of Gaussian noise at inference (ablation) | CO-affected model · CIFAR-10 ResNet18 | AutoAttack accuracy (%): 41.64 -> 43.53 | Improved |
| increase uniform noise range from 10/255 to 16/255 (ablation) | CO-affected model · CIFAR-10 ResNet18 | AutoAttack accuracy (%): 23.27 -> 43.53 | Improved |
| train with noise-augmented data instead of relying on CO (ablation) | noise-augmented models · CIFAR-10 ResNet18 | adversarial robustness | Hurt |
| induce CO with L_co on top 10 percent channels (ablation) | FGSM-MEP train 8/255 · CIFAR-10 ResNet18 | CO occurrence | Improved |
| use proposed feature-difference regularizer instead of ConvergeSmooth at 12/255 | stable FAT · CIFAR-10 ResNet18 | AutoAttack accuracy (%): 19.11 -> 22.08 | Improved |
| compare high-variance channel masking against feature-difference masking (ablation) | CO-affected model · CIFAR-10 ResNet18 | clean and adversarial accuracy | Improved |
Closest papers by meaning
- Layer-Aware Analysis of Catastrophic Overfitting: Revealing the Pseudo-Robust Shortcut Dependency (ICML 2024 · similarity 0.92)
- Overfitting in adversarially robust deep learning (ICML 2020 · similarity 0.91)
- Taxonomy Driven Fast Adversarial Training (AAAI 2024 · similarity 0.88)
- Efficient local linearity regularization to overcome catastrophic overfitting (ICLR 2024 · similarity 0.88)
- Adversarial Training Can Provably Improve Robustness: Theoretical Analysis of Feature Learning Process Under Structured Data (ICLR 2025 · similarity 0.88)
- Benign Overfitting in Adversarial Training of Neural Networks (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.
