Research map / Adversarial robustness and security / Certified and empirical robustness
Neural Architecture Search for Wide Spectrum Adversarial Robustness
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| search with 6 adversarial noise strengths instead of 1 (ablation) | Wsr-NAS on CIFAR-10, TRADES-1 · CIFAR-10 | average robust accuracy (%): 21.7% -> 22.9% | Helped |
| search with 11 adversarial noise strengths instead of 6 (ablation) | Wsr-NAS on CIFAR-10, TRADES-1 · CIFAR-10 | average robust accuracy (%): 22.4% -> 24.2% | Helped |
| use more adversarial noise strengths during search (ablation) | Wsr-NAS · CIFAR-10 search | search time: 0.08 days -> 0.48 days per adversarial noise strength increase | Mixed |
| add the AN-Estimator to adversarial noise generation during search (ablation) | Wsr-NAS · CIFAR-10 search iteration | adversarial example generation time (s): 1151.5 -> 576.3 | Helped |
| add the EWSS instead of calculating architecture gradients over adversarial validation sets (ablation) | Wsr-NAS · CIFAR-10 search iteration | search time (s): 191.3 -> 37.1 | Helped |
| add attention to E-ANE (ablation) | E-ANE adversarial noise estimation · CIFAR-10 test set | average robust accuracy (%): 71.1% -> 48.9% | Helped |
| use A-ANE instead of E-ANE (ablation) | adversarial noise estimation · CIFAR-10 test set | average robust accuracy (%): 48.9% -> 43.6% | Helped |
| replace learned adversarial noise estimation with weighted averaging (ablation) | AN-Estimator · CIFAR-10 test set | average robust accuracy (%): 38.7% -> 67.4% | Hurt |
| use WsrNet-Basic rather than RACL | ImageNet models trained with Fast Adversarial Training · ImageNet | average robust accuracy (%): 8.5 -> 9.0 | Helped |
| use WsrNet-Basic rather than a ResNet-50 | CIFAR-10 models trained with TRADES-1 · CIFAR-10 | average robust accuracy (%): 14.8% -> 22.4% | Helped |
| train one model with multiple noise strengths rather than use WsrNet-Basic | ResNet-50 versus WsrNet-Basic · CIFAR-10 | average robust accuracy (%): 22.4% -> 14.8% | Hurt |
| use WsrNet-Basic rather than WRN-34-10 | TRADES-1 models on CIFAR-10 · CIFAR-10 | natural accuracy (%): 88.6% -> 90.4% | Helped |
Closest papers by meaning
- Robust NAS under adversarial training: benchmark, theory, and beyond (ICLR 2024 · similarity 0.92)
- Towards Accurate and Robust Architectures via Neural Architecture Search (CVPR 2024 · similarity 0.91)
- DSRNA: Differentiable Search of Robust Neural Architectures (CVPR 2021 · similarity 0.91)
- First Line of Defense: A Robust First Layer Mitigates Adversarial Attacks (AAAI 2025 · similarity 0.90)
- A Combination of Noise and Bilateral Filters Achieve Supralinear and Scalable Adversarial Robustness in CNNs (CVPR 2026 · similarity 0.90)
- 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.