Research map / Adversarial robustness and security / Black-box and transfer attacks
Boosting Generative Adversarial Transferability with Self-supervised Vision Transformer Features
Shangbo Wu, Yu-an Tan, Ruinan Ma, Wencong Ma, Dehua Zhu, Yuanzhang Li
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 |
|---|---|---|---|
| jointly train on DINO (CL) and MAE (MIM) features instead of a single self-supervised ViT (ablation) | dSVA · ImageNet black-box transfer (12 target models) | fooling rate: +13.70% on average over single-model variants | Improved |
| use self-supervised ViT features (DINO or MAE) instead of supervised ViT (ablation) | dSVA · ImageNet black-box transfer | fooling rate | Improved |
| use facet-level query/key features instead of whole-layer token output (ablation) | dSVA (DINO, MAE) · ImageNet black-box transfer | fooling rate | Improved |
| add attention saliency regularization to supervised ViT surrogate (ablation) | dSVA with supervised ViT · ImageNet black-box transfer | fooling rate | Hurt |
| add attention saliency regularization to DINO and MAE variants (ablation) | dSVA (DINO), dSVA (MAE) · ImageNet black-box transfer | fooling rate | Improved |
| add attention saliency regularization to the joint variant (ablation) | dSVA (Joint) · ImageNet black-box transfer | fooling rate | Mixed |
| move feature layer from earlier layers deeper, up to layer 10 (ablation) | dSVA · ImageNet black-box transfer | fooling rate | Improved |
| use final layer 11 instead of penultimate layer 10 (ablation) | dSVA · ImageNet black-box transfer | fooling rate | Hurt |
| set lambda to 0.5 rather than 0.9 (more MIM weight) (ablation) | dSVA (Joint) · ImageNet black-box transfer | fooling rate | Improved |
| dSVA (Joint) versus prior attacks on adversarially trained models | dSVA (Joint) · ImageNet defense models | fooling rate: +32.98% average margin over compared attacks | Improved |
| dSVA versus BIA and CDA on cross-domain targets | dSVA · coarse and fine-grained cross-domain classification | fooling rate: approximately 6% on average | Improved |
| use ConvNet surrogates (CDA/BIA) versus dSVA on same-architecture targets | VGG-16, Res-50, Den-121 · ImageNet black-box transfer | fooling rate: VGG-16: 99.31 vs 96.78 | Hurt |
Closest papers by meaning
- A Theoretical Analysis of Self-Supervised Learning for Vision Transformers (ICLR 2025 · similarity 0.89)
- Improving the Adversarial Transferability of Vision Transformers with Virtual Dense Connection (AAAI 2024 · similarity 0.88)
- Boosting the Transferability of Adversarial Attack on Vision Transformer with Adaptive Token Tuning (NeurIPS 2024 · similarity 0.88)
- Harnessing the Computation Redundancy in ViTs to Boost Adversarial Transferability (NeurIPS 2025 · similarity 0.88)
- Vision Transformers Need Registers (ICLR 2024 · similarity 0.88)
- Attacking Transformers with Feature Diversity Adversarial Perturbation (AAAI 2024 · similarity 0.88)
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.
