Research map / Adversarial robustness and security / Certified and empirical robustness
Exploring the Relationship Between Architectural Design and Adversarially Robust Generalization
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 |
|---|---|---|---|
| Use PVTv2 or CoAtNet rather than CNN architectures for adversarial training | Adversarially trained architectures · CIFAR-10 | Worst-case Acc: 33.54 (PVTv2), 32.17 (CoAtNet) vs 14.32 (ResNet) | Helped |
| Use ViT rather than ResNet | Adversarially trained architectures · ImageNet | Worst-Case Acc: 24.77 vs 20.08 | Helped |
| Add attention to more CoAtNet stages rather than use convolution throughout (ablation) | CoAtNet on CIFAR-10 · CIFAR-10 | Worst-case: 27.64 (CCCC) -> 30.30 (CTTT) | Helped |
| Replace attention in the final stage with convolution (CTTT to CTTC) (ablation) | CoAtNet on CIFAR-10 · CIFAR-10 | Worst-case: 30.30 -> 29.96 | Hurt |
| Replace attention with convolution in the third stage (CTTC to CTCC) (ablation) | CoAtNet on CIFAR-10 · CIFAR-10 | Worst-case: 29.96 -> 26.69 | Hurt |
| Reduce PVTv2 patch size from 4 to 2 (ablation) | PVTv2 · CIFAR-10 | Worst-case Acc: 33.54 -> 31.16 | Hurt |
| Reduce PVTv2 patch size from 2 to 1 (ablation) | PVTv2 · CIFAR-10 | Worst-case Acc: 31.16 -> 21.10 | Hurt |
| Reduce ViT patch size from 4 to 2 (ablation) | ViT · CIFAR-10 | Worst-case Acc: 29.24 -> 23.78 | Hurt |
| Reduce patch size (ablation) | ViT and PVTv2 · CIFAR-10 | PGD-ℓ∞ accuracy: PVTv2: 46.48 -> 54.59; ViT: 42.77 -> 45.45 | Helped |
| Reduce patch size (ablation) | ViT and PVTv2 · CIFAR-10 | PGD-ℓ1 and PGD-ℓ2 accuracy: PVTv2 PGD-ℓ2: 35.77 -> 23.47; PGD-ℓ1: 46.14 -> 24.76 | Hurt |
| Add ℓ1 regularization during adversarial training (ablation) | ResNet · CIFAR-10 | Worst-case robust generalization: +2% | Helped |
| Increase ℓ1 regularization strength (ablation) | ResNet · CIFAR-10 | Clean accuracy | Hurt |
Closest papers by meaning
- Fantastic Robustness Measures: The Secrets of Robust Generalization (NeurIPS 2023 · similarity 0.91)
- Exploring Architectural Ingredients of Adversarially Robust Deep Neural Networks (NeurIPS 2021 · similarity 0.91)
- Revisiting Adversarial Training for ImageNet: Architectures, Training and Generalization across Threat Models (NeurIPS 2023 · similarity 0.91)
- Revisiting Residual Networks for Adversarial Robustness (CVPR 2023 · similarity 0.91)
- When NAS Meets Robustness: In Search of Robust Architectures Against Adversarial Attacks (CVPR 2020 · similarity 0.90)
- Robust NAS under adversarial training: benchmark, theory, and beyond (ICLR 2024 · similarity 0.90)
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.