1330-GRM Context-Aware Privacy Detection in Vision-Language Models: Evaluating Input Strategies and Model Architectures
Vision-Language Models (VLMs) support visual assistance by answering questions about images, but these images may contain sensitive information. The privacy risk depends on the image information and the user's question request. This study investigates question-dependent privacy detection and different input and prompting strategies. Preliminary experiments across four VLMs achieve 54%-84% accuracy. Initial question-based experiments show that question-only input achieves higher accuracy than direct image–question input. We evaluate structured and sequential privacy-aware processing to improve context-dependent privacy recognition.
Authors
- Ramisa Fariha Joyee
Publication Details
- Journal
- DigitalCommons - Kennesaw State University (Kennesaw State University)
- Published
- 2026-11-20
- Primary Topic
- Privacy-Preserving Technologies in Data
- Type
- article
- Field-Weighted Citation Impact
- 0.00