Research map / Adversarial robustness and security / Attacks on text and graphs
Boosting Transferability in Vision-Language Attacks via Diversification along the Intersection Region of Adversarial Trajectory
Sensen Gao, Xiaojun Jia, Xuhong Ren, Ivor Tsang, Qing Guo
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 image augmentation to SGA (ablation) | SGA · Flickr30K ITR, ALBEF to TCL | TR R@1 ASR: 70.07 -> 87.88 | Improved |
| sample from intersection region of adversarial trajectory with text-guided selection (vs SGA) | SGA · Flickr30K ITR, ALBEF to CLIP_ViT | TR R@1 ASR: 36.69 -> 46.26 | Improved |
| use proposed method instead of SGA on same-architecture transfer | SGA · Flickr30K ITR, ALBEF to TCL | TR R@1 ASR: 87.88 -> 91.57 | Improved |
| use proposed method instead of SGA in white-box setting | SGA · Flickr30K ITR, ALBEF white-box | TR R@1 ASR: 99.9 -> 99.9 | No clear effect |
| use proposed method instead of SGA for cross-architecture transfer to CLIP_CNN | SGA · Flickr30K ITR, CLIP_ViT to CLIP_CNN | TR R@1 ASR: 53.26 -> 64.88 | Improved |
| increase clean-image weight lambda in adversarial text generation (ablation) | adversarial text loss · Flickr30K ITR, ALBEF to CLIP_ViT | TR R@1 ASR: 42.21 (0.2) -> 46.26 (0.6) -> 46.13 (0.8) | Mixed |
| increase number of intersection-region samples N beyond 5 (ablation) | proposed attack · Flickr30K ITR, ALBEF to CLIP_ViT | TR R@1 ASR: 46.26 (N=5) -> 45.4 (N=7) | No clear effect |
| use proposed attack instead of SGA for cross-task transfer to captioning | SGA · ITR to MSCOCO image captioning | CIDEr (lower is stronger attack): 91.4 -> 88.3 | Improved |
| add intersection-region diversification and text deviation | attack time cost · Flickr30K ITR | time per adversarial example: 1.58-2.28 times longer than SGA | Hurt |
| remove diversification along intersection region vs remove text deviation (ablation) | proposed attack · Flickr30K ITR, ALBEF to other models | ASR | Mixed |
Closest papers by meaning
- Harnessing the Computation Redundancy in ViTs to Boost Adversarial Transferability (NeurIPS 2025 · similarity 0.90)
- Chain of Attack: On the Robustness of Vision-Language Models Against Transfer-Based Adversarial Attacks (CVPR 2025 · similarity 0.90)
- Failures to Find Transferable Image Jailbreaks Between Vision-Language Models (ICLR 2025 · similarity 0.89)
- One Prompt Word is Enough to Boost Adversarial Robustness for Pre-trained Vision-Language Models (CVPR 2024 · similarity 0.89)
- Enhancing Adversarial Transferability with Adversarial Weight Tuning (AAAI 2025 · similarity 0.89)
- Attention! Your Vision Language Model Could Be Maliciously Manipulated (NeurIPS 2025 · 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.
