Research map / Adversarial robustness and security / Certified and empirical robustness
Fast Adversarial Training with Smooth Convergence
Mengnan 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 batch-based ConvergeSmooth to FGSM-MEP | FGSM-MEP · CIFAR-10 (eps=16/255) | AutoAttack accuracy: 18.98 -> 23.68 | Improved |
| add batch-based ConvergeSmooth to FGSM-RS | FGSM-RS · CIFAR-10 (eps=16/255) | AutoAttack accuracy: 14.84 -> 19.43 | Improved |
| add batch-based ConvergeSmooth to FGSM-MEP | FGSM-MEP · CIFAR-10 (eps=16/255) | PGD-50 accuracy: 26.56 -> 32.95 | Improved |
| add weight centralization to FGSM-RS | FGSM-RS · CIFAR-10 (eps=16/255) | AutoAttack accuracy: 14.84 -> 17.14 | Improved |
| add example-based ConvergeSmooth to FGSM-MEP | FGSM-MEP · CIFAR-10 (eps=16/255) | AutoAttack accuracy: 18.98 -> 17.74 | Hurt |
| add batch-based ConvergeSmooth to FGSM-BP on Tiny ImageNet | FGSM-BP · Tiny ImageNet (PreActResNet18) | AutoAttack accuracy: 3.62 -> 5.95 | Improved |
| add ConvergeSmooth to FGSM-MEP with WideResNet34-10 | FGSM-MEP · CIFAR-10 (WideResNet34-10) | AutoAttack accuracy: 15.43 -> 24.27 | Improved |
| increase gamma_max above 0.09 (ablation) | B-RS · CIFAR-100 (ResNet18) | stability | Hurt |
| use dynamic convergence stride (gamma_max/gamma_min = 1.5) instead of static (ablation) | B-RS · CIFAR-100 (ResNet18) | robustness | Improved |
| increase loss coefficient w1 from 0 to 0.9 (ablation) | B-RS · CIFAR-100 (ResNet18, eps=12/255) | AutoAttack accuracy: 15.28 -> 13.53 | Hurt |
| increase w2 from 0.5 to 1.0 (ablation) | B-RS · CIFAR-100 (ResNet18, eps=12/255) | AutoAttack accuracy: 12.23 -> 15.28 | Improved |
| set weight centralization coefficient w3 to 0 vs 0.1 (ablation) | weight centralization · CIFAR-10 / CIFAR-100 | stability | Improved |
Closest papers by meaning
- Preventing Catastrophic Overfitting in Fast Adversarial Training: A Bi-level Optimization Perspective (ECCV 2024 · similarity 0.92)
- Improving Fast Adversarial Training With Prior-Guided Knowledge (TPAMI 2024 · similarity 0.91)
- Quadratic Upper Bound for Boosting Robustness (ICML 2025 · similarity 0.89)
- Adversarial Initialization with Universal Adversarial Perturbation: A New Approach to Fast Adversarial Training (AAAI 2024 · similarity 0.87)
- Uniformly Stable Algorithms for Adversarial Training and Beyond (ICML 2024 · similarity 0.87)
- Efficient local linearity regularization to overcome catastrophic overfitting (ICLR 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.
