Research map / Adversarial robustness and security / Certified and empirical robustness
Robust Network Architecture Search via Feature Distortion Restraining
Yaguan Qian, Shenghui Huang, Bin Wang, Xiang Ling, Xiaohui Guan, Zhaoquan Gu, Shaoning Zeng, Wujie Zhou, Haijiang 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 network-vulnerability constraint (RNAS-H vs DARTS search) | DARTS · CIFAR-10 (PGD AT) | PGD20 robust accuracy (%): 44.91 -> 53.48 | Improved |
| add network-vulnerability constraint (RNAS-H vs DARTS search) | DARTS · CIFAR-10 (PGD AT) | AutoAttack robust accuracy (%): 39.98 -> 50.54 | Improved |
| tighten vulnerability bound from H=0.0001 to H=0.00001 (RNAS-L) (ablation) | RNAS · CIFAR-10 (PGD AT) | PGD20 robust accuracy (%): 53.48 -> 54.85 | Improved |
| tighten vulnerability bound from H=0.0001 to H=0.00001 (RNAS-L) (ablation) | RNAS · CIFAR-100 (transferred) | AutoAttack robust accuracy (%): 28.29 -> 27.37 | Hurt |
| tighten vulnerability bound from H=0.0001 to H=0.00001 (RNAS-L) (ablation) | RNAS · SVHN (transferred) | AutoAttack robust accuracy (%): 77.26 -> 75.89 | Hurt |
| tighten vulnerability bound from H=0.0001 to H=0.00001 (RNAS-L) (ablation) | RNAS · Tiny-ImageNet (transferred) | AutoAttack robust accuracy (%): 24.22 -> 23.02 | Hurt |
| tighten vulnerability bound from H=0.0001 to H=0.00001 (RNAS-L) (ablation) | RNAS · Tiny-ImageNet (searched directly) | AutoAttack robust accuracy (%): 24.82 -> 25.74 | Improved |
| decrease H (stronger constraint) toward very small values (ablation) | RNAS · CIFAR-10/CIFAR-100/SVHN (Fig. 5) | PGD20 robust accuracy | Mixed |
| search architecture directly on Tiny-ImageNet instead of transferring from CIFAR-10 (RNAS-L) (ablation) | RNAS-L · Tiny-ImageNet | AutoAttack robust accuracy (%): 23.02 -> 25.74 | Improved |
| use RNAS cell with only standard training | RNAS-L · CIFAR-10 (standard training) | FGSM accuracy (%): 54.72 -> 65.32 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- AdvRush — AdvRush: Searching for Adversarially Robust Neural Architectures (ICCV 2021) · better on 8 of 8 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- DSRNA: Differentiable Search of Robust Neural Architectures (CVPR 2021 · similarity 0.92)
- Exploring Architectural Ingredients of Adversarially Robust Deep Neural Networks (NeurIPS 2021 · similarity 0.92)
- Neural Architecture Design and Robustness: A Dataset (ICLR 2023 · similarity 0.91)
- Adversarially Robust Neural Architecture Search for Graph Neural Networks (CVPR 2023 · similarity 0.91)
- Towards Accurate and Robust Architectures via Neural Architecture Search (CVPR 2024 · similarity 0.91)
- Robust NAS under adversarial training: benchmark, theory, and beyond (ICLR 2024 · similarity 0.91)
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.
