Research map / Adversarial robustness and security / Certified and empirical robustness
Understanding and Improving Fast Adversarial Training
Maksym Andriushchenko, Nicolas Flammarion
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 |
|---|---|---|---|
| reduce FGSM step size to 7/255 (no random init) (ablation) | FGSM AT · CIFAR-10 eps=8/255 | PGD-50-10 adversarial accuracy (%): 36.35 -> 45.35 | Improved |
| add random initialization step to FGSM (ablation) | FGSM AT · CIFAR-10 eps=8/255 | PGD-50-10 adversarial accuracy (%): 36.35 -> 45.60 | Improved |
| add random step to FGSM at larger eps (ablation) | FGSM AT · CIFAR-10 eps=10/255 | adversarial accuracy (%): 0% | Hurt |
| project PGD-10 perturbations onto corners of l_inf ball (ablation) | PGD-10 AT · CIFAR-10 eps=8/255 | adversarial accuracy (%): 50.48 -> 50.64 | No clear effect |
| zero out the Laplace-like filter w4 (ablation) | Single-layer CNN after catastrophic overfitting · CIFAR-10 | gradient alignment: 0.08 -> 0.71 | Improved |
| add GradAlign regularizer to FGSM AT | FGSM AT · CIFAR-10 and SVHN, large eps | PGD-50-10 adversarial accuracy | Improved |
| add GradAlign regularizer to PGD-2 AT | PGD-2 AT · CIFAR-10 eps=16/255 | PGD-50-10 adversarial accuracy | Improved |
| add GradAlign regularizer to FGSM AT | FGSM AT · CIFAR-10 and SVHN, small eps | adversarial accuracy | No clear effect |
| add GradAlign to FGSM AT | FGSM AT · ImageNet eps 2/255 and 4/255 | adversarial accuracy | No clear effect |
| add GradAlign to FGSM AT | FGSM AT · Training time | training cost: 3x slowdown | Hurt |
Closest papers by meaning
- Understanding Catastrophic Overfitting in Single-step Adversarial Training (AAAI 2021 · similarity 0.92)
- Adversarial Initialization with Universal Adversarial Perturbation: A New Approach to Fast Adversarial Training (AAAI 2024 · similarity 0.90)
- Efficient local linearity regularization to overcome catastrophic overfitting (ICLR 2024 · similarity 0.88)
- Taxonomy Driven Fast Adversarial Training (AAAI 2024 · similarity 0.87)
- Benign Overfitting in Adversarial Training of Neural Networks (ICML 2024 · similarity 0.87)
- Layer-Aware Analysis of Catastrophic Overfitting: Revealing the Pseudo-Robust Shortcut Dependency (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.
