Research map / Adversarial robustness and security / Certified and empirical robustness
Towards Efficient Adversarial Training on Vision Transformers
Boxi Wu, Jindong Gu, Zhifeng Li, Deng Cai, Xiaofei He, Wei Liu
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 |
|---|---|---|---|
| random patch dropping (40%) added to Fast AT (ablation) | Fast AT ViT-Base · ImageNet eps=2/255 | PGD-100 robust accuracy: 50.06 -> 36.81 | Hurt |
| attention-guided dropping (AGAT) added to Fast AT (ablation) | Fast AT ViT-Base · ImageNet eps=2/255 | PGD-100 robust accuracy: 50.06 -> 51.11 | Improved |
| attention-guided dropping (AGAT) added to Fast AT (ablation) | Fast AT ViT-Base · ImageNet | training time: 86H -> 56H | Improved |
| attention-guided dropping (AGAT) added to Fast AT (ablation) | Fast AT CaiT-S36 · ImageNet | training time: 76H -> 53H | Improved |
| attention-guided dropping (AGAT) added to Fast AT (ablation) | Fast AT LeViT-384 · ImageNet eps=2/255 | PGD-100 robust accuracy: 46.13 -> 48.59 | Improved |
| attention-guided dropping (AGAT) added to Fast AT (ablation) | Fast AT ViT-Small · ImageNet eps=2/255 | PGD-100 robust accuracy: 49.41 -> 48.85 | No clear effect |
| attention-guided instead of random dropping at 60% computation (ablation) | Random dropping ViT-Tiny · ImageNet eps=2/255 | PGD robust accuracy: 21.01 -> 46.60 | Improved |
| attention-guided instead of random dropping at 40% computation (ablation) | Random dropping ViT-Base · ImageNet eps=2/255 | PGD robust accuracy: 36.81 -> 50.02 | Improved |
| increase AGAT dropped computation from 40% to 60% (ablation) | AGAT ViT-Base · ImageNet eps=4/255 | PGD robust accuracy: 34.78 -> 32.90 | Hurt |
| random patch dropping added to Fast AT (ablation) | Fast AT ViT-Tiny · ImageNet eps=2/255 | PGD-100 robust accuracy: 48.28 -> 31.16 | Hurt |
| evaluate AGAT under APGD-CE instead of PGD (ablation) | AGAT ViT-Base · ImageNet eps=2/255 | robust accuracy: 50.01 -> 51.23 | No clear effect |
| add attention dropout to Fast AT | Fast AT ViTs · ImageNet | training stability | Improved |
Closest papers by meaning
- Random Entangled Tokens for Adversarially Robust Vision Transformer (CVPR 2024 · similarity 0.90)
- Towards Understanding and Improving Adversarial Robustness of Vision Transformers (CVPR 2024 · similarity 0.90)
- Harnessing the Computation Redundancy in ViTs to Boost Adversarial Transferability (NeurIPS 2025 · similarity 0.89)
- Improving the Adversarial Transferability of Vision Transformers with Virtual Dense Connection (AAAI 2024 · similarity 0.89)
- Boosting the Transferability of Adversarial Attack on Vision Transformer with Adaptive Token Tuning (NeurIPS 2024 · similarity 0.89)
- You Only Need Less Attention at Each Stage in Vision Transformers (CVPR 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.
