Research map / Adversarial robustness and security / Certified and empirical robustness
Trade-Off Between Robustness and Accuracy of Vision Transformers
Yanxi Li, Chang Xu
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 |
|---|---|---|---|
| Replace single-adapter / head-only tuning with two adapters plus patch-level gated fusion (TORA-ViT) (ablation) | ViT-B/16 frozen backbone · ImageNet-1K clean and PGD (λ=0.1) | top-1 accuracy: clean 84.1 vs 82.5 (single adapter); PGD 23.3 vs 15.1 | Improved |
| Replace adapter-level AdapterFusion with patch-level trade-off-aware gated fusion (ablation) | TORA-ViT · ImageNet-1K PGD (λ=0.9) | top-1 accuracy: 57.5 vs 55.3 | Improved |
| Tune only a new classification head on frozen ViT (ablation) | ViT-B/16 · ImageNet-1K PGD (λ=0.1) | top-1 accuracy: 15.5 vs 23.3 for full model | Hurt |
| Increase λ from 0.1 to 0.5 (more robustness weight) | TORA-ViT-B/16 · ImageNet-1K clean vs PGD | top-1 accuracy: clean 84.1→83.7 (-0.4); PGD 23.3→38.0 (+14.7) | Mixed |
| Increase λ from 0.5 to 0.9 | TORA-ViT-B/16 · ImageNet-A | top-1 accuracy: 39.2 → 22.2 (-17.0) | Hurt |
| Increase λ from 0.5 to 0.9 | TORA-ViT-B/16 · ImageNet-1K FGSM and PGD | top-1 accuracy: FGSM 54.7→74.2; PGD 38.0→57.5 | Improved |
| Joint prediction of accuracy and robustness heads vs. single head at λ=0.5 (ablation) | TORA-ViT-B/16 · ImageNet-1K clean | top-1 accuracy: joint 83.66 vs accuracy head 83.38 and robustness head 83.01; joint is lower-than-acc-head under PGD (37.99 vs 39.78) | Mixed |
Closest papers by meaning
- Vision Transformers Are Robust Learners (AAAI 2022 · similarity 0.92)
- Understanding Robustness of Transformers for Image Classification (ICCV 2021 · similarity 0.92)
- Towards Robust Vision Transformer (CVPR 2022 · similarity 0.91)
- Vision Transformers Beat WideResNets on Small Scale Datasets Adversarial Robustness (AAAI 2025 · similarity 0.89)
- Random Entangled Tokens for Adversarially Robust Vision Transformer (CVPR 2024 · similarity 0.87)
- Improving Interpretation Faithfulness for Vision Transformers (ICML 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.
