Research map / Adversarial robustness and security / Attacks on text and graphs
Prefill-Time Interventions against Adversarial Attacks on Large Vision-Language Models
Zhewen Yao, Yao Zhu, Shiliang Zhang
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 S_inc to S_anom (ablation) | PTI · In-domain COCO, Qwen2.5-VL-3B | AUC: 0.971 -> 0.979 | Improved |
| add S_inc to S_anom (ablation) | PTI · In-domain COCO, Qwen2.5-VL-3B | TPR@5%FPR: 87.64 -> 91.30 | Improved |
| add S_anom to S_inc (ablation) | PTI · OOD ImageNet, Qwen2.5-VL-3B | FPR@95%TPR: 29.08 -> 25.18 | Improved |
| add S_inc to S_anom (ablation) | PTI · OOD ImageNet, Qwen2.5-VL-3B | AUC: 0.918 -> 0.936 | Improved |
| add S_inc to S_anom (ablation) | PTI · Stress test, Qwen2.5-VL-3B | FPR@95%TPR: 30.34 -> 20.62 | Improved |
| add S_anom to S_inc (ablation) | PTI · In-domain COCO, Qwen2.5-VL-3B | AUC: 0.924 -> 0.979 | Improved |
| add self-distillation warm-up to S_inc (ablation) | S_inc · In-domain COCO, Qwen2.5-VL-3B | TPR@5%FPR: 65.83 -> 70.19 | Improved |
| add self-distillation warm-up to S_inc (ablation) | S_inc · Stress test, Qwen2.5-VL-3B | FPR@95%TPR: 26.96 -> 22.14 | Improved |
| add self-distillation warm-up to S_inc (ablation) | S_inc · OOD ImageNet, Qwen2.5-VL-3B | AUC: 0.891 -> 0.911 | Improved |
| replace PIP with PTI | detector · In-domain, 4 LVLMs avg | AUC: 0.762 -> 0.980 | Improved |
| replace MirrorCheck with PTI | detector · In-domain, 4 LVLMs avg | time overhead: 1.1809 -> 0.0931 | Improved |
| use deeper vision-tower layers for anomaly estimator (ablation) | S_anom · Layer choice, 3 LVLMs | detection accuracy | Hurt |
Closest papers by meaning
- A Frustratingly Simple Yet Highly Effective Attack Baseline: Over 90% Success Rate Against the Strong Black-box Models of GPT-4.5/4o/o1 (NeurIPS 2025 · similarity 0.90)
- Pandora's Box: Towards Building Universal Attackers against Real-World Large Vision-Language Models (NeurIPS 2024 · similarity 0.90)
- Stage-wise Attention-Guided Region Sequencing for Adversarial Attacks on Large Vision-Language Models (NeurIPS 2026 · similarity 0.89)
- R-TPT: Improving Adversarial Robustness of Vision-Language Models through Test-Time Prompt Tuning (CVPR 2025 · similarity 0.89)
- Attention! Your Vision Language Model Could Be Maliciously Manipulated (NeurIPS 2025 · similarity 0.89)
- SLADE: Shielding against Dual Exploits in Large Vision-Language Models (CVPR 2025 · 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.
