Research map / Adversarial robustness and security / Certified and empirical robustness
Exploring Architectural Ingredients of Adversarially Robust Deep Neural Networks
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 |
|---|---|---|---|
| reduce Stage-3 depth from 5 to 1 while deepening Stages 1-2 (d9-7-1) (ablation) | WRN-34-10 · CIFAR-10 SAT | PGD20 accuracy (%): 51.43 -> 54.03 | Helped |
| reduce Stage-3 width from 10 to 4 (w10-10-4) (ablation) | WRN-34-10 · CIFAR-10 SAT | PGD20 accuracy (%): 52.04 -> 54.28 | Helped |
| reduce Stage-3 width from 10 to 2 (ablation) | WRN-34-10 · CIFAR-10 SAT | PGD20 accuracy (%): 52.04 -> 53.95 | Helped |
| remove Stage-3 entirely (d5-5-0) (ablation) | WRN-34-10 · CIFAR-10 SAT | PGD20 accuracy (%): about 2% lower than baseline | Hurt |
| reduce depth/width at Stages 1 or 2 (ablation) | WRN-34-10 · CIFAR-10 SAT | PGD20 accuracy (%) | Hurt |
| combine top depth configs with w10-10-4 (ablation) | WRN-34-10 · CIFAR-10 SAT | PGD20 accuracy (%): about 54% in all cases | No clear effect |
| scale w10-10-4 by gamma from 1.0 to 2.0 (ablation) | w10-10-4 · CIFAR-10 SAT | PGD20 accuracy (%): 54.99 -> 55.35 | Helped |
| scale w10-10-4 down to gamma 0.25 (ablation) | w10-10-4 · CIFAR-10 SAT | PGD20 accuracy (%): 54.99 -> 50.90 | Hurt |
| apply robust width reconfiguration (WRN-34-R vs WRN-34-12) | WRN-34 with TRADES · CIFAR-10 | AutoAttack last (%): 48.48 -> 50.90 | Helped |
| apply robust width reconfiguration (WRN-34-R vs WRN-34-12) | WRN-34 with SAT · CIFAR-10 | AutoAttack last (%): 46.06 -> 46.75 | Helped |
| apply reconfiguration (-R) to VGG-11, DN-121, DARTS | VGG/DN/DARTS with SAT · CIFAR-10 | AutoAttack last (%): 37.21 -> 38.44 (VGG-11) | Helped |
| apply WRN-34-R reconfiguration under CW attack | WRN-34 with SAT/MART · CIFAR-100 | CW-inf last (%): 43.78 -> 43.52 (SAT) | Hurt |
Closest papers by meaning
- Revisiting Residual Networks for Adversarial Robustness (CVPR 2023 · similarity 0.93)
- When NAS Meets Robustness: In Search of Robust Architectures Against Adversarial Attacks (CVPR 2020 · similarity 0.92)
- Adversarial Training Can Provably Improve Robustness: Theoretical Analysis of Feature Learning Process Under Structured Data (ICLR 2025 · similarity 0.90)
- Towards Accurate and Robust Architectures via Neural Architecture Search (CVPR 2024 · similarity 0.90)
- Robust NAS under adversarial training: benchmark, theory, and beyond (ICLR 2024 · similarity 0.90)
- Interpreting Robustness Proofs of Deep Neural Networks (ICLR 2024 · 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.