Research map / Adversarial robustness and security / Certified and empirical robustness
Adversarial Weight Perturbation Helps Robust Generalization
Dongxian Wu, Shu-Tao Xia, Yisen Wang
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 adversarial weight perturbation to AT | AT, WideResNet-34-10 · CIFAR-10, L∞ | PGD-20 robustness: 56.10 -> 58.14 | Improved |
| Add adversarial weight perturbation to TRADES | TRADES, WideResNet-34-10 · CIFAR-10, L∞ | PGD-20 robustness: 56.33 -> 59.27 | Improved |
| Add adversarial weight perturbation to MART | MART, WideResNet-34-10 · CIFAR-10, L∞ | PGD-20 robustness: 58.56 -> 60.68 | Improved |
| Add adversarial weight perturbation to RST | RST, WideResNet-28-10 · CIFAR-10, L∞ | PGD-20 robustness: 62.60 -> 63.73 | Improved |
| Add adversarial weight perturbation to RST | RST, WideResNet-28-10 · CIFAR-10, L∞ | natural accuracy: 89.69 -> 88.25 | Hurt |
| Use more than one alternation between input and weight perturbation (ablation) | AT-AWP · CIFAR-10 | test robustness | No clear effect |
| Increase weight-perturbation optimization steps K2 from 1 to 5 or 10 (ablation) | AT-AWP · CIFAR-10 | test robustness | No clear effect |
| Use L1 or L2/Frobenius norm for weight-perturbation size (ablation) | AT-AWP · CIFAR-10 | test robustness | No clear effect |
| Set weight perturbation size too large (γ = 2×10−2) (ablation) | AT-AWP · CIFAR-10 | test robustness | Hurt |
| Replace adversarial weight perturbation with random weight perturbation at γ ≤ 0.3 (ablation) | AT-RWP · CIFAR-10 | test robustness | No clear effect |
| Add L1 regularization to AT (ablation) | AT · CIFAR-10, L∞ | best robustness: 52.79 -> 51.95 | Hurt |
| Add L2 regularization to AT (ablation) | AT · CIFAR-10, L∞ | best robustness: 52.79 -> 51.60 | Hurt |
Closest papers by meaning
- Relating Adversarially Robust Generalization to Flat Minima (ICCV 2021 · similarity 0.92)
- Formalizing Generalization and Adversarial Robustness of Neural Networks to Weight Perturbations (NeurIPS 2021 · similarity 0.89)
- Revisiting the Trade-Off Between Accuracy and Robustness via Weight Distribution of Filters (TPAMI 2024 · similarity 0.89)
- Adversarial Weight Perturbation Improves Generalization in Graph Neural Networks (AAAI 2023 · similarity 0.89)
- Adversarial Training Can Provably Improve Robustness: Theoretical Analysis of Feature Learning Process Under Structured Data (ICLR 2025 · similarity 0.87)
- Enhancing Adversarial Transferability with Adversarial Weight Tuning (AAAI 2025 · 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.
