Research map / Adversarial robustness and security / Attacks on text and graphs
DDGA: Dirichlet Distributional Gradient Aggregation for Transferable Vision-Language Adversarial Attacks
Yiwei You, Jiaan Wei, Zan Chen, Bo Wang
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 |
|---|---|---|---|
| replace the learned Dirichlet policy with uniform alpha (ablation) | DDGA · Flickr30K, CLIP-ViT -> ALBEF TR R@1 | attack success rate (%): 38.21 -> 41.19 | Improved |
| add covariance-guided orthogonal perturbation (ablation) | DDGA · Flickr30K, CLIP-ViT -> ALBEF TR R@1 | attack success rate (%): 39.14 -> 41.19 | Improved |
| add adversarial text attack (ablation) | DDGA · Flickr30K, CLIP-ViT -> ALBEF TR R@1 | attack success rate (%): 36.77 -> 41.19 | Improved |
| add covariance-guided perturbation (ablation) | DDGA · Flickr30K, CLIP-ViT -> CLIP-CNN IR R@1 | attack success rate (%): 75.01 -> 75.13 | No clear effect |
| add learned Dirichlet policy (ablation) | DDGA · Flickr30K, CLIP-ViT -> CLIP-CNN IR R@1 | attack success rate (%): 74.80 -> 75.13 | No clear effect |
| use DDGA instead of SA-AET | adversarial attack on CLIP-ViT source · MSCOCO, ALBEF TR R@1 | attack success rate (%): 57.64 -> 66.46 | Improved |
| use DDGA instead of SA-AET | adversarial attack on CLIP-ViT source · Flickr30K, CLIP-CNN target TR R@1 | attack success rate (%): 71.01 -> 70.24 | Hurt |
| set covariance injection factor lambda too small or too large (ablation) | DDGA · Flickr30K ITR | attack success rate | Hurt |
| use DDGA instead of SA-AET for cross-task transfer to captioning | adversarial attack · Flickr30K, BLIP captioning | CIDEr: 35.14 -> 28.83 | Improved |
Closest papers by meaning
- Semantic-Aligned Adversarial Evolution Triangle for High-Transferability Vision-Language Attack (TPAMI 2025 · similarity 0.89)
- Boosting Transferability in Vision-Language Attacks via Diversification along the Intersection Region of Adversarial Trajectory (ECCV 2024 · similarity 0.88)
- Chain of Attack: On the Robustness of Vision-Language Models Against Transfer-Based Adversarial Attacks (CVPR 2025 · similarity 0.88)
- Anyattack: Towards Large-scale Self-supervised Adversarial Attacks on Vision-language Models (CVPR 2025 · similarity 0.87)
- Exploring Visual Vulnerabilities via Multi-Loss Adversarial Search for Jailbreaking Vision-Language Models (CVPR 2025 · similarity 0.86)
- R-TPT: Improving Adversarial Robustness of Vision-Language Models through Test-Time Prompt Tuning (CVPR 2025 · similarity 0.86)
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.
