Research map / Vision-language and multimodal / Cross-modal learning
MADA-Attack: Transferable Multi-modal Attention Distraction Adversarial Attack against Vision Language Models
Zhihan Qin, Jiahao Chen, Chunyi Zhou, Yuwen Pu, Chunqiang Hu, Xiaolei Liu, Shouling Ji
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 Semantic Token Manipulation (structured text perturbation) instead of random token selection (ablation) | MADA-Attack · MSCOCO captioning, BLIP2 | CIDEr ASR (%): 45.63 -> 49.42 | Improved |
| add Fused Embedding Training (joint text and image optimization) instead of image-only (ablation) | MADA-Attack · MSCOCO captioning, BLIP2 | CIDEr ASR (%): 38.99 -> 49.42 | Improved |
| add Adaptive Data Augmentation (low early probability, increasing later) instead of fixed probability (ablation) | MADA-Attack · MSCOCO captioning, BLIP2 | CIDEr ASR (%): 36.51 -> 49.42 | Improved |
| add text-guided embedding-level optimization to UAP for zero-shot classification | XTransfer Base · ZS classification average (BLIP2) | ASR (%): 70.83 -> 82.60 | Improved |
| MADA-Attack vs best baseline on BLIP2 captioning | XTransfer Base · Flickr30K captioning, BLIP2 | CIDEr ASR (%): 33.3 -> 44.39 | Improved |
| use MADA-Attack on commercial closed-source models | C-PGC and XTransfer Naive · GPT-4o, Gemini, Claude captioning | CIDEr ASR (%): GPT-4o 74.50 -> 94.59 | Improved |
| MADA-Attack on VizWiz VQA vs best baseline | MiniGPT4 VQA · VizWiz | ASR (%) | No clear effect |
| MADA-Attack vs XTransfer Base on LLaVA-7B captioning CIDEr | XTransfer Base · Flickr30K captioning, LLaVA-7B | CIDEr ASR (%): 97.05 -> 98.25 | No clear effect |
| increase image perturbation budget from 4/255 to 12/255 (ablation) | MADA-Attack · BLIP2 captioning | caption disruption (qualitative) | Improved |
| MADA-Attack training time vs other UAP methods | UAP training · ZS classification | training time | No clear effect |
Closest papers by meaning
- Mutual-Modality Adversarial Attack with Semantic Perturbation (AAAI 2024 · similarity 0.88)
- MoDA: Modulation Adapter for Fine-Grained Visual Understanding in Instructional MLLMs (ICML 2026 · similarity 0.88)
- HQA-VLAttack: Towards High Quality Adversarial Attack on Vision-Language Pre-Trained Models (NeurIPS 2025 · similarity 0.87)
- Stage-wise Attention-Guided Region Sequencing for Adversarial Attacks on Large Vision-Language Models (NeurIPS 2026 · similarity 0.87)
- VERA-V: Variational Inference Framework for Jailbreaking Vision-Language Models (ICML 2026 · similarity 0.87)
- ETA: Evaluating Then Aligning Safety of Vision Language Models at Inference Time (ICLR 2025 · similarity 0.87)
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.
