Research map / Adversarial robustness and security / Black-box and transfer attacks
On the Robustness of Vision Transformers to Adversarial Examples
Kaleel Mahmood, Rigel Mahmood, Marten van Dijk
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 |
|---|---|---|---|
| use a ViT instead of a CNN or BiT model | single classifier · CIFAR-10/ImageNet white-box | robust accuracy under PGD and MIM: less than 6% | No clear effect |
| use a ViT instead of a CNN | single classifier · CIFAR-10/ImageNet white-box C&W and APGD | robust accuracy: 0% | No clear effect |
| evaluate adversarial examples crafted on BiT-50x1 against ViT-L-16 | BiT-M-50x1 adversarial examples · CIFAR-10, CIFAR-100, ImageNet transferability | transferability: 5.7%, 15.5% and 11.8% | Hurt |
| evaluate adversarial examples crafted on ViT-L-16 against BiT-M-50x1 | ViT-L-16 adversarial examples · CIFAR-10, CIFAR-100, ImageNet transferability | transferability: 42.5%, 47.6% and 34.3% | Hurt |
| tune SAGA coefficients toward the ViT gradient instead of equal 0.5/0.5 weights (ablation) | SAGA · CIFAR-10 ViT/BiT ensemble | average robust accuracy: 47.5% -> 26.0% | Improved |
| tune SAGA coefficients instead of equal 0.5/0.5 weights (ablation) | SAGA · CIFAR-100 ViT/BiT ensemble | average robust accuracy: 26.3% -> 15.6% | Improved |
| tune SAGA coefficients instead of equal 0.5/0.5 weights (ablation) | SAGA · ImageNet ViT/BiT ensemble | average robust accuracy: 28.5% -> 12.7% | Improved |
| attack a ViT+BiT random-selection ensemble with SAGA | ensemble defense · CIFAR-10, CIFAR-100, ImageNet white-box | attack success rate: 74.0%, 84.4% and 91.8% | Hurt |
| use a ViT+BiT random-selection ensemble instead of a single model against RayS | ensemble defense · CIFAR-10, CIFAR-100, ImageNet RayS black-box | robust accuracy: +66.5%, +73% and +32.1% | Improved |
| use a ViT+BiT random-selection ensemble instead of a single model against the Adaptive attack | ensemble defense · CIFAR-10 Adaptive black-box | robust accuracy: 57%, +26.2% | Improved |
| use a ViT+BiT ensemble instead of single classifiers | ensemble defense · clean accuracy | clean accuracy: 98.2%, 92.83% and 85.37% | No clear effect |
Closest papers by meaning
- Are Transformers more robust than CNNs? (NeurIPS 2021 · similarity 0.92)
- Towards Transferable Adversarial Attacks on Vision Transformers (AAAI 2022 · similarity 0.92)
- Towards Understanding and Improving Adversarial Robustness of Vision Transformers (CVPR 2024 · similarity 0.91)
- Attacking Transformers with Feature Diversity Adversarial Perturbation (AAAI 2024 · similarity 0.90)
- Harnessing the Computation Redundancy in ViTs to Boost Adversarial Transferability (NeurIPS 2025 · similarity 0.90)
- Boosting the Transferability of Adversarial Attack on Vision Transformer with Adaptive Token Tuning (NeurIPS 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.
