Research map / Adversarial robustness and security / Certified and empirical robustness
Do Wider Neural Networks Really Help Adversarial Robustness?
Boxi Wu, Jinghui Chen, Deng Cai, Xiaofei He, Quanquan Gu
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 |
|---|---|---|---|
| increase TRADES lambda from 6 to 18 (ablation) | TRADES WRN-34-10 · CIFAR10 PGD-20 | robust accuracy (%): 54.18 -> 56.43 | Improved |
| increase TRADES lambda from 6 to 9 (ablation) | TRADES WRN-34-1 · CIFAR10 PGD-20 | robust accuracy (%): 47.81 -> 48.01 | Improved |
| increase TRADES lambda from 9 to 21 (ablation) | TRADES WRN-34-1 · CIFAR10 PGD-20 | robust accuracy (%): 48.01 -> 46.26 | Hurt |
| increase TRADES lambda from 6 to 21 (ablation) | TRADES WRN-34-10 · CIFAR10 | natural accuracy (%): 84.90 -> 81.74 | Hurt |
| increase TRADES lambda from 6 to 21 (ablation) | TRADES WRN-34-10 · CIFAR10 | perturbation stability (%): 67.25 -> 72.11 | Improved |
| increase network width from 1 to 10 at lambda=6 (ablation) | TRADES WRN-34 · CIFAR10 | perturbation stability (%): 69.33 -> 67.25 | Hurt |
| increase network width from 1 to 10 at lambda=6 | TRADES WRN-34 · CIFAR10 PGD-20 | robust accuracy (%): 47.81 -> 54.18 | Improved |
| increase network width from 5 to 10 at lambda=6 (ablation) | TRADES WRN-34 · CIFAR10 PGD-20 | robust accuracy (%): 54.45 -> 54.18 | No clear effect |
| increase lambda from 1.00 to 2.00 for adversarial training (ablation) | Adversarial training WRN-34-10 · CIFAR10 PGD-20 | robust accuracy (%): 50.12 -> 55.78 | Improved |
| increase lambda from 6 to 18 (ablation) | TRADES WRN-34-10 · CIFAR100 PGD | robust accuracy (%): 30.48 -> 32.98 | Improved |
| increase lambda from 6 to 12 (ablation) | TRADES DenseNet-BC-40 k=12 · CIFAR10 PGD | robust accuracy (%): 44.79 -> 44.66 | No clear effect |
| increase lambda from 6 to 12 (ablation) | TRADES DenseNet-BC-40 k=64 · CIFAR10 PGD | robust accuracy (%): 55.51 -> 55.85 | Improved |
Closest papers by meaning
- Exploring Architectural Ingredients of Adversarially Robust Deep Neural Networks (NeurIPS 2021 · similarity 0.91)
- Towards Understanding the Regularization of Adversarial Robustness on Neural Networks (ICML 2020 · similarity 0.89)
- l∞-Robustness and Beyond: Unleashing Efficient Adversarial Training (ECCV 2022 · similarity 0.89)
- Theoretical Analysis of Robust Overfitting for Wide DNNs: An NTK Approach (ICLR 2024 · similarity 0.88)
- Adversarial Robustness Limits via Scaling-Law and Human-Alignment Studies (ICML 2024 · similarity 0.88)
- Adversarial Training Can Provably Improve Robustness: Theoretical Analysis of Feature Learning Process Under Structured Data (ICLR 2025 · similarity 0.88)
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.
