Research map / Adversarial robustness and security / Certified and empirical robustness
Make Some Noise: Reliable and Efficient Single-Step Adversarial Training
Pau de Jorge, Adel Bibi, Riccardo Volpi, Amartya Sanyal, Philip Torr, Grégory Rogez, Puneet K. Dokania
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 strong noise augmentation and remove clipping around clean sample (N-FGSM vs RS-FGSM) | FGSM adversarial training · ImageNet eps=6/255 | PGD50-10 accuracy (%): 16.49 -> 17.12 | Improved |
| add noise augmentation (N-FGSM) to FGSM | FGSM · ImageNet eps=6/255 | PGD50-10 accuracy (%): 0.08 -> 17.12 | Improved |
| clip perturbation around the clean sample (ablation) | N-FGSM · CIFAR-10 eps=8/255 | PGD50-10 accuracy | Hurt |
| increase noise magnitude k beyond eps (ablation) | N-FGSM (no clipping) · CIFAR-10 large eps | avoids catastrophic overfitting | Improved |
| increase step size alpha from 1.25eps to 1.5eps keeping k=1eps (ablation) | N-FGSM · CIFAR-10 eps=8/255 | PGD50-10 accuracy (%): 49.68% -> 0 | Hurt |
| add noise to FGSM step (vs larger alpha without noise) (ablation) | FGSM · CIFAR-10 eps=8/255 | PGD50-10 accuracy | Improved |
| increase training eps from 8/255 to 12/255 (ablation) | RS-FGSM · CIFAR-10 eps=8/255 test | PGD50-10 accuracy (%): 46.08 -> 0.0 | Hurt |
| increase training eps from 8/255 to 12/255 (ablation) | GradAlign · CIFAR-10 eps=8/255 test | PGD50-10 accuracy (%): 48.14 -> 50.6 robust; 81.9 -> 73.29 clean | Mixed |
| use long training schedule with final model instead of fast schedule (ablation) | N-FGSM · CIFAR-10 eps=8/255 | PGD50-10 accuracy (%): 48.12 -> 36.56 | Hurt |
| add early stopping (best model) to long schedule (ablation) | N-FGSM · CIFAR-10 eps=8/255 | PGD50-10 accuracy (%): 36.56 -> 48.48 | Improved |
| use N-FGSM instead of GradAlign | single-step AT · CIFAR-10/100, SVHN PreActResNet18 | robust accuracy and cost: 3x speed-up | Improved |
Closest papers by meaning
- Understanding and Improving Fast Adversarial Training (NeurIPS 2020 · similarity 0.91)
- Understanding Catastrophic Overfitting in Single-step Adversarial Training (AAAI 2021 · similarity 0.90)
- Preventing Catastrophic Overfitting in Fast Adversarial Training: A Bi-level Optimization Perspective (ECCV 2024 · similarity 0.89)
- Adversarial Initialization with Universal Adversarial Perturbation: A New Approach to Fast Adversarial Training (AAAI 2024 · similarity 0.87)
- Benign Overfitting in Adversarial Training of Neural Networks (ICML 2024 · similarity 0.86)
- Efficient local linearity regularization to overcome catastrophic overfitting (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.
