ImgSan: Utility-Based Image Privacy Sanitizer for Online Visual Question Answering

Summary

ImgSan is an image privacy sanitization approach for online Visual Question Answering that applies region-specific obfuscation with adaptive intensity. It scores the semantic relevance of image regions to the question, then strengthens protection for regions that are both relevant and sensitive while keeping answers accurate. On PrivacyVQA2K, a benchmark of 1,000 images with 2,002 annotated privacy-sensitive objects that the authors constructed, it reportedly balances privacy protection and response utility better than existing baselines.