VSF-Med: A Vulnerability Scoring Framework for Medical Vision-Language Models
Published in IEEE ISBI 2026, 2026
| Project Page | arXiv |
VSF-Med scores vulnerabilities in medical Vision-Language Models (VLMs) using text-prompt attack templates, imperceptible visual perturbations, and an eight-dimensional risk rubric. We evaluated 68,478 attack scenarios across five current models.
Recommended citation: Sadanandan, B., & Behzadan, V. (2026). VSF-Med: A Vulnerability Scoring Framework for Medical Vision-Language Models. IEEE ISBI 2026.
Download Paper