Research map / Adversarial robustness and security / Certified and empirical robustness
AdvRush: Searching for Adversarially Robust Neural Architectures
Jisoo Mok, Byunggook Na, Hyeokjun Choe, Sungroh Yoon
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 the input-loss-curvature (Hessian Frobenius norm) regularizer to architecture search | DARTS supernet · CIFAR-10 (adv. trained) | AutoAttack robust accuracy: 46.79% -> 50.05% | Improved |
| search with AdvRush instead of PDARTS, standard-trained evaluation | PDARTS · CIFAR-10 (standard trained) | FGSM robust accuracy: 54.51% -> 55.91% | Improved |
| use AdvRush architecture instead of RobNet-free, adversarially trained | RobNet-free · CIFAR-10 (adv. trained) | AutoAttack robust accuracy: 48.56% -> 50.05% | Improved |
| transfer the AdvRush architecture instead of PDARTS | PDARTS · CIFAR-100 | PGD-20 robust accuracy: 25.83% -> 30.15% | Improved |
| transfer the AdvRush architecture instead of PDARTS | PDARTS · SVHN | PGD-20 robust accuracy: 89.58% -> 91.14% | Improved |
| transfer the AdvRush architecture instead of PDARTS | PDARTS · Tiny-ImageNet | PGD-20 robust accuracy: 22.74% -> 23.58% | Improved |
| lower gamma from 0.01 to 0.001 (ablation) | AdvRush · CIFAR-10 (adv. trained) | clean accuracy: 87.30% -> 85.65% | Hurt |
| raise gamma from 0.01 to 0.1 (ablation) | AdvRush · CIFAR-10 (adv. trained) | clean accuracy: 87.30% -> 83.03% | Hurt |
| raise gamma from 0.01 to 0.1 (ablation) | AdvRush · CIFAR-10 (adv. trained) | PGD-20 robust accuracy: 53.07% -> 53.67% | No clear effect |
| replace the Hessian regularizer with FGSM adversarial loss during supernet search (no warm-up, 50 adv epochs) (ablation) | AdvRush · CIFAR-10 (adv. trained) | PGD-20 robust accuracy: 53.07% -> 48.76% | Hurt |
| replace the Hessian regularizer with PGD adversarial loss during supernet search (no warm-up, 50 adv epochs) (ablation) | AdvRush · CIFAR-10 (adv. trained) | PGD-20 robust accuracy: 53.07% -> 46.37% | Hurt |
| replace the Hessian regularizer with FGSM adversarial loss during supernet search (50 warm-up, 10 adv epochs) (ablation) | AdvRush · CIFAR-10 (adv. trained) | clean accuracy: 87.30% -> 82.78% | Hurt |
Later papers that compared with this method
3 later accepted papers report AdvRush (7@96) as a baseline in their results tables; 3 of them report a better number on most of the results they share with it.
- Revisiting Residual Networks for Adversarial Robustness (CVPR 2023 · better on 2 of 2 shared results)
- Generalizable Lightweight Proxy for Robust NAS against Diverse Perturbations (NeurIPS 2023 · better on 2 of 2 shared results)
- Robust Network Architecture Search via Feature Distortion Restraining (ECCV 2022 · 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
- Robust NAS under adversarial training: benchmark, theory, and beyond (ICLR 2024 · similarity 0.91)
- An Introduction to Adversarially Robust Deep Learning (TPAMI 2023 · similarity 0.90)
- Towards Deep Learning Models Resistant to Adversarial Attacks (ICLR 2018 · similarity 0.90)
- Robust Network Architecture Search via Feature Distortion Restraining (ECCV 2022 · similarity 0.89)
- Towards Accurate and Robust Architectures via Neural Architecture Search (CVPR 2024 · similarity 0.89)
- Adversarial Training Can Provably Improve Robustness: Theoretical Analysis of Feature Learning Process Under Structured Data (ICLR 2025 · similarity 0.89)
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.
