Research map / Adversarial robustness and security / Attacks on text and graphs
Misaligned Roles, Misplaced Images: Structural Input Perturbations Expose Multimodal Alignment Blind Spots
Erfan Shayegani, G M Shahariar, Sara Abdali, Lei Yu, Nael Abu-Ghazaleh, Yue Dong
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 |
|---|---|---|---|
| swap user and assistant roles in the prompt (Role Confusion) (ablation) | Phi-3.5-vision · AdvBench | ASR (%) target-string: 6.35 -> 65.96 | Hurt |
| swap user and assistant roles in the prompt (Role Confusion) (ablation) | LLaVA-1.5-7B · AdvBench | ASR (%) target-string: 22.12 -> 78.46 | Hurt |
| compose role swap with image-at-end modality manipulation (ablation) | Qwen2-VL-7B · AdvBench | ASR (%) target-string: 8.08 / 5.96 -> 32.88 | Hurt |
| move image token to the beginning of the assistant turn (ablation) | Qwen2-VL-7B · AdvBench | ASR (%) target-string: 0.58 -> 37.31 | Hurt |
| compose role swap with AIM jailbreak (ablation) | Qwen2-VL-7B · AdvBench | ASR (%) target-string: 35.38 -> 96.54 | Hurt |
| compose role swap with Evil Confidant jailbreak (ablation) | Qwen2-VL-7B · AdvBench | ASR (%) target-string: 18.27 -> 1.35 | Improved |
| add adversarial training on RMA-perturbed harmful and benign prompts | LLaVA-1.5-7B · AdvBench (avg 8 settings) | ASR (%): 75.04 -> 2.60 | Improved |
| add adversarial training on RMA-perturbed harmful and benign prompts | Qwen2-VL-7B · HarmBench (avg 8 settings) | ASR (%): 31.64 -> 0.00 | Improved |
| add adversarial training on RMA-perturbed prompts | LLaVA-1.5-7B · VQA-V2 | accuracy (%): 67.88 -> 63.07 | Hurt |
| add adversarial training on RMA-perturbed prompts | Phi-3.5-vision · VQA-V2 | accuracy (%): 62.88 -> 71.92 | Improved |
| add adversarial training on RMA-perturbed prompts | Qwen2-VL-7B · VQA-V2 | accuracy (%): 81.53 -> 81.73 | No clear effect |
| add adversarial training on RMA-perturbed prompts | LLaVA-1.5-7B · Alpaca benign prompts | refusal rate (%): 0.77 -> 6.73 (img out_swap) | Hurt |
Closest papers by meaning
- ETA: Evaluating Then Aligning Safety of Vision Language Models at Inference Time (ICLR 2025 · similarity 0.89)
- Jailbreak in pieces: Compositional Adversarial Attacks on Multi-Modal Language Models (ICLR 2024 · similarity 0.89)
- Do We Really Need Curated Malicious Data for Safety Alignment in Multi-modal Large Language Models? (CVPR 2025 · similarity 0.89)
- DeepAlign: Mitigating Modality Conflict through Modality-Specific Alignment (CVPR 2026 · similarity 0.87)
- GuardAlign: Robust Safety Alignment in Multimodal Large Language Models (ICLR 2026 · similarity 0.87)
- Enhance Modality Robustness in Text-Centric Multimodal Alignment with Adversarial Prompting (AAAI 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.
