Vision Language Model Helps Private Information De-Identification in Vision Data
Researchers have developed VisShield, a new framework to enhance privacy in Vision Language Models (VLMs). This framework includes a specialized dataset called OPTIC, designed for instruction tuning, and a tailored training methodology. VisShield aims to accurately locate sensitive text within images and apply privacy protection, outperforming existing methods in experiments. AI
IMPACT This research could enable more secure applications of vision-language models, particularly in sensitive domains like healthcare.