Research map / Adversarial robustness and security / Attacks on text and graphs
Contextual Image Attack: How Visual Context Exposes Multimodal Safety Vulnerabilities
Yuan Xiong, Miao Ziqi, Lijun Li, Chen Qian, Jie Li, Jing Shao
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 |
|---|---|---|---|
| embed harmful query in a generated contextual image (CIA) vs best prior baseline VisCo | CIA · MMSafetyBench-tiny | ASR (%): 85.71 -> 86.31 (GPT-4o); 88.10 -> 91.07 (Qwen2.5-VL-72B) | Improved |
| CIA vs VisCo on other models in MMSafetyBench-tiny | CIA · MMSafetyBench-tiny (Gemini-2.0, InternVL2.5) | ASR (%): 92.86 -> 85.12 (Gemini-2.0); 92.86 -> 82.14 (InternVL2.5) | Hurt |
| CIA vs FigStep and other baselines on SafeBench-tiny average | CIA · SafeBench-tiny | ASR (%): 83.20 -> 90.80 vs VisCo; 40.80 FigStep | Improved |
| add frame structure (ablation) | CIA · SafeBench-tiny | ASR (%): 38 -> 60 | Improved |
| add toxicity obfuscation of visual text (ablation) | CIA · SafeBench-tiny | ASR (%): 36 -> 60 | Improved |
| use visual input and image context instead of text-only or blank image (ablation) | CIA · SafeBench-tiny | ASR (%) | Improved |
| add semantic alignment refinement and contextual elements over initial image (ablation) | CIA · SafeBench-tiny | ASR (%) | Improved |
| add emoji insertion augmentation (ablation) | CIA (Qwen2.5-VL-72B) · SafeBench-tiny by strategy | ASR (%): 66.00 -> 78.00 (Demonstration); 60.00 -> 30.00 (Structured Content) | Mixed |
| add noise injection augmentation (ablation) | CIA (Qwen2.5-VL-72B) · SafeBench-tiny by strategy | ASR (%): 66.00 -> 36.00 (Demonstration); 46.00 -> 54.00 (Sequential Path) | Mixed |
| add safety icon augmentation (ablation) | CIA (Qwen2.5-VL-72B) · SafeBench-tiny by strategy | ASR (%): 46.00 -> 62.00 (Sequential Path); 44.00 -> 36.00 (Dialogue Layout) | Mixed |
| choose Demonstration vs Dialogue Layout visualization strategy (ablation) | CIA · SafeBench-tiny (average over 5 models) | ASR (%): 56.40 -> 90.80 | Improved |
Closest papers by meaning
- Jailbreaking Vision-Language Models Through the Visual Modality (ICML 2026 · similarity 0.90)
- Exploring Visual Vulnerabilities via Multi-Loss Adversarial Search for Jailbreaking Vision-Language Models (CVPR 2025 · similarity 0.88)
- When the Prompt Becomes Visual: Vision-Centric Jailbreak Attacks for Large Image Editing Models (ICML 2026 · similarity 0.88)
- FigStep: Jailbreaking Large Vision-Language Models via Typographic Visual Prompts (AAAI 2025 · similarity 0.88)
- Distraction is All You Need for Multimodal Large Language Model Jailbreaking (CVPR 2025 · similarity 0.88)
- MIDAS: Multi-Image Dispersion and Semantic Reconstruction for Jailbreaking MLLMs (ICLR 2026 · 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.
