Research map / Adversarial robustness and security / Attacks on text and graphs
AdaShield: Safeguarding Multimodal Large Language Models from Structure-based Attack via Adaptive Shield Prompting
Yu Wang, Xiaogeng Liu, Yu Li, Muhao Chen, Chaowei Xiao
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 |
|---|---|---|---|
| prepend adaptive auto-refined defense prompt (AdaShield-A) vs vanilla | LLaVA 1.5-13B · QR attack | ASR: 75.75 -> 15.22 | Improved |
| prepend AdaShield-A defense prompt vs vanilla | CogVLM-chat-v1.1 · FigStep attack | ASR: 85.19 -> 0.00 | Improved |
| use AdaShield-A instead of static prompt AdaShield-S (ablation) | LLaVA 1.5-13B · QR attack | ASR: 24.43 -> 15.22 | Improved |
| add post-hoc harm detector and detoxifier (MLLMP) vs vanilla | LLaVA 1.5-13B · QR attack | ASR: 75.75 -> 77.96 | No clear effect |
| static defense prompt FSD on benign tasks | CogVLM-chat-v1.1 · MM-Vet benign | Total: 50.0 -> 27.4 | Hurt |
| AdaShield-S on MiniGPT-v2 benign tasks (over-defense) | MiniGPT-v2-13B · MM-Vet benign | Total: 14.8 -> 1.4 | Hurt |
| retrieve defense prompts by similarity (threshold) instead of random choice (ablation) | AdaShield-A · QR attack, LLaVA 1.5-13B | ASR: 18.20 -> 15.22 | Improved |
| add instruction to execute benign queries safely (intuition 4) (ablation) | AdaShield-S · MM-Vet, CogVLM-chat-v1.1 | Total: 14.3 -> 45.9 | Improved |
| specify explicit refusal response ('I am sorry') in defense prompt (ablation) | AdaShield-S · QR and FigStep | ASR | Improved |
| train defense pool on Hard scenarios only vs All (ablation) | AdaShield-A · QR (All test) | ASR: 15.22 -> 14.63 | Improved |
| transfer defense prompts learned with another MLLM (ablation) | AdaShield-A · QR, CogVLM-chat-v1.1 | ASR: 16.07 -> 7.33 | Improved |
Closest papers by meaning
- SafePTR: Token-Level Jailbreak Defense in Multimodal LLMs via Prune-then-Restore Mechanism (NeurIPS 2025 · similarity 0.90)
- MIDAS: Multi-Image Dispersion and Semantic Reconstruction for Jailbreaking MLLMs (ICLR 2026 · similarity 0.90)
- MMJ-Bench: A Comprehensive Study on Jailbreak Attacks and Defenses for Vision Language Models (AAAI 2025 · similarity 0.90)
- Distraction is All You Need for Multimodal Large Language Model Jailbreaking (CVPR 2025 · similarity 0.89)
- Immune: Improving Safety Against Jailbreaks in Multi-modal LLMs via Inference-Time Alignment (CVPR 2025 · similarity 0.89)
- Many-shot Jailbreaking (NeurIPS 2024 · similarity 0.89)
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.
