Research map / Adversarial robustness and security / Certified and empirical robustness
Pyramid Adversarial Training Improves ViT Performance
Check 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 |
|---|---|---|---|
| Replace pixel-wise adversarial perturbations with multi-scale pyramid perturbations | ViT-B/16 with RegViT training · ImageNet-1K val | top-1 accuracy: 80.42 -> 81.71 | Helped |
| Replace pixel-wise adversarial perturbations with multi-scale pyramid perturbations | ViT-B/16 with RegViT training · ImageNet-C | mCE: 47.68 -> 44.99 | Helped |
| Replace pixel-wise adversarial perturbations with multi-scale pyramid perturbations | ViT-B/16 with RegViT training · ImageNet-Rendition | top-1 accuracy: 45.39 -> 47.66 | Helped |
| Replace pixel-wise adversarial perturbations with multi-scale pyramid perturbations | ViT-B/16 with RegViT training · ImageNet-Sketch | top-1 accuracy: 34.40 -> 36.77 | Helped |
| Add matched Dropout and stochastic depth to the adversarial branch (ablation) | PyramidAT on ViT-B/16 · ImageNet-1K val | top-1 accuracy: 79.43 -> 81.71 | Helped |
| Add matched Dropout and stochastic depth to the adversarial branch (ablation) | PyramidAT on ViT-B/16 · ImageNet-C | mCE: 54.70 -> 44.99 | Helped |
| Remove Dropout from the adversarial update (ablation) | PyramidAT on ViT-B/16 · ImageNet-Rendition | top-1 accuracy: 47.66 -> 52.34 | Helped |
| Use a 2-level rather than a 3-level pyramid (ablation) | PyramidAT on ViT-B/16 · ImageNet-1K val | top-1 accuracy: 81.65 vs 81.71 | Hurt |
| Use a 4-level rather than a 3-level pyramid (ablation) | PyramidAT on ViT-B/16 · ImageNet-1K val | top-1 accuracy: 81.71 -> 81.66 | No clear effect |
| Apply PixelAT with RandAugment magnitude 0.4 instead of the clean baseline (ablation) | ViT-Ti/16 · ImageNet-1K val | top-1 accuracy: 64.27 -> 62.78 | Hurt |
| Apply PyramidAT with RandAugment magnitude 0.4 instead of the clean baseline (ablation) | ViT-Ti/16 · ImageNet-1K val | top-1 accuracy: 64.27 -> 65.61 | Helped |
| Increase PixelAT attack steps from 1 to 3 (ablation) | ViT-Ti/16, RAm=0.1 · ImageNet-1K val | top-1 accuracy: 64.66 -> 59.91 | Hurt |
Closest papers by meaning
- When Adversarial Training Meets Vision Transformers: Recipes from Training to Architecture (NeurIPS 2022 · similarity 0.90)
- Vision Transformers Beat WideResNets on Small Scale Datasets Adversarial Robustness (AAAI 2025 · similarity 0.88)
- GradQ-ViT: Robust and Efficient Gradient Quantization for Vision Transformers (AAAI 2025 · similarity 0.88)
- Bi-ViT: Pushing the Limit of Vision Transformer Quantization (AAAI 2024 · similarity 0.87)
- Random Entangled Tokens for Adversarially Robust Vision Transformer (CVPR 2024 · similarity 0.87)
- Improving the Adversarial Transferability of Vision Transformers with Virtual Dense Connection (AAAI 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.