Research map / Adversarial robustness and security / Black-box and transfer attacks
Transferable Adversarial Attack for Both Vision Transformers and Convolutional Networks via Momentum Integrated Gradients
Wenshuo Ma, Yidong Li, Xiaofeng Jia, Wei Xu
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 |
|---|---|---|---|
| add integrated gradients and momentum to the iterative attack (ablation) | MIG · ImageNet 1k subset (target ViT-L, source DeiT-S) | attack success rate (%): 40.23 -> 86.33 (+46.10) | Improved |
| add integrated gradients and momentum to the iterative attack (ablation) | MIG · ImageNet 1k subset (target ViT-L, source DeiT-B) | attack success rate (%): 35.63 -> 87.85 (+52.22) | Improved |
| add integrated gradients and momentum to the iterative attack (ablation) | MIG · ImageNet 1k subset (target ViT-L, source MNAS) | attack success rate (%): 16.77 -> 37.45 (+20.68) | Improved |
| add integrated gradients and momentum to the iterative attack (ablation) | MIG · ImageNet 1k subset (target ViT-L, source Incep-v4) | attack success rate (%): 15.75 -> 50.25 (+34.50) | Improved |
| use IG with a single-step update only (ablation) | MIG · ImageNet 1k subset (target ViT-L) | attack success rate (%): poor performance | Hurt |
| add momentum on top of IG and iteration (ablation) | MIG · ImageNet 1k subset (target ViT-L) | attack success rate (%): 8.62% ~ 24.14% improvement | Improved |
| use logit or IG ensemble of two source models vs perturbation ensemble (ablation) | MIG · ImageNet 5k val | mean attack success rate (%): 3.87% ~ 25.63% over the two single models | Improved |
| ensemble perturbations directly (ablation) | MIG · ImageNet 5k val (VGG19+DeiT-S) | mean attack success rate (%): 81.71 vs 91.33 for logit ensemble (ViT MASR 87.63 vs 92.22) | Hurt |
| combine MIG with DIM, TIM and SIM (ablation) | DIM/TIM/SIM · ImageNet 5k val (source DeiT-S) | attack success rate (%): about 5% ~ 23% gain | Improved |
| raise momentum decay factor above 1 (ablation) | MIG · ImageNet 5k val (source DeiT-S) | attack success rate (%): optimal at mu = 1, then decreases | Hurt |
| use MIG instead of FGSM (DeiT-B source, CNN targets) | MIG · ImageNet 5k val | mean attack success rate (%): +47.77% | Improved |
| use MIG instead of MI-RE (DeiT-B source, CNN targets) | MIG · ImageNet 5k val | mean attack success rate (%): +10.93% | Improved |
Closest papers by meaning
- Boosting the Transferability of Adversarial Attack on Vision Transformer with Adaptive Token Tuning (NeurIPS 2024 · similarity 0.91)
- Improving the Adversarial Transferability of Vision Transformers with Virtual Dense Connection (AAAI 2024 · similarity 0.90)
- Harnessing the Computation Redundancy in ViTs to Boost Adversarial Transferability (NeurIPS 2025 · similarity 0.90)
- Improving Adversarial Transferability on Vision Transformers via Forward Propagation Refinement (CVPR 2025 · similarity 0.89)
- Enhancing Adversarial Transferability with Adversarial Weight Tuning (AAAI 2025 · similarity 0.89)
- Towards Understanding and Improving Adversarial Robustness of 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.
