Research map / Adversarial robustness and security / Certified and empirical robustness
Neural Architecture Dilation for 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 |
|---|---|---|---|
| Add optimized architecture dilation to a WRN32-10 backbone | WRN32-10 · CIFAR-10 | PGD-20 accuracy: 45.84 -> 53.43 | Helped |
| Add FLOPs-constrained dilation instead of unconstrained dilation (ablation) | NADAR on WRN32-10 · CIFAR-10 | PGD-20 accuracy: 52.84 -> 53.43 | Hurt |
| Add FLOPs-constrained dilation instead of unconstrained dilation (ablation) | NADAR on WRN32-10 · CIFAR-100 | PGD-20 accuracy: 27.77 -> 28.40 | Helped |
| Add architecture dilation to ResNet-50 | ResNet-50 · Tiny-ImageNet | PGD-20 accuracy: 22.16 -> 22.70 | Helped |
| Add separate optimization objectives for backbone and dilation rather than training the hybrid jointly (ablation) | NADAR on WRN32-10 · CIFAR-10 | PGD-20 accuracy: 45.97$\pm$0.18 -> 48.53$\pm$0.33 | Helped |
| Add the standard performance constraint (ablation) | NADAR with separate objectives · CIFAR-10 | natural accuracy: 84.79$\pm$0.55 -> 85.97$\pm$0.26 | Helped |
| Add the standard performance constraint (ablation) | NADAR with separate objectives · CIFAR-10 | PGD-20 accuracy: 48.53$\pm$0.33 -> 53.18$\pm$0.25 | Helped |
| Increase dilation from 3 × 3 to 3 × 6 cells (ablation) | NADAR on WRN32-10 · CIFAR-10 | PGD-20 accuracy: 47.78$\pm$0.41 -> 49.63$\pm$0.18 | Helped |
| Increase dilation from 3 × 6 to 3 × 9 cells (ablation) | NADAR on WRN32-10 · CIFAR-10 | PGD-20 accuracy: 49.63$\pm$0.18 -> 45.25$\pm$0.57 | Hurt |
| Increase dilation from 3 × 6 to 3 × 9 cells (ablation) | NADAR on WRN32-10 · CIFAR-10 | natural accuracy: 86.28$\pm$0.26 -> 85.63$\pm$0.12 | Hurt |
| Add the FLOPs constraint (ablation) | NADAR on ResNet-18 · MNIST | PGD-40 accuracy: 98.97 -> 98.90 | Hurt |
| Add the FLOPs constraint (ablation) | NADAR on ResNet-18 · MNIST-M | PGD-40 accuracy: 92.62 -> 91.79 | Hurt |
Closest papers by meaning
- Exploring Architectural Ingredients of Adversarially Robust Deep Neural Networks (NeurIPS 2021 · similarity 0.91)
- Neural Architecture Search for Wide Spectrum Adversarial Robustness (AAAI 2023 · similarity 0.90)
- Robust Models are less Over-Confident (NeurIPS 2022 · similarity 0.89)
- Adversarial Training Can Provably Improve Robustness: Theoretical Analysis of Feature Learning Process Under Structured Data (ICLR 2025 · similarity 0.88)
- Towards Accurate and Robust Architectures via Neural Architecture Search (CVPR 2024 · similarity 0.88)
- Robust NAS under adversarial training: benchmark, theory, and beyond (ICLR 2024 · similarity 0.87)
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.