Mitigating inferential disclosure in wearable VLM interactions

Develop privacy-preserving methods for vision–language model interactions on wearable cameras that prevent inferential disclosure of sensitive attributes—such as location identification from architectural details, signage, and geographical features—especially when query targets contain identifying features.

Background

Even when explicit identifiers are hidden or cropped, vision–LLMs can infer sensitive attributes from background cues, leading to privacy risks for both wearers and bystanders.

The authors highlight that such inferential disclosure persists in real-world use and remains an unresolved obstacle for privacy-preserving interaction designs on wearable vision systems.