ScreenRecompose: Semantic-Aware Generative Content Replacement for Privacy Protection in Screen Sharing
Screen sharing in online conferencing can unintentionally expose private information on the presenter’s screen. Existing protections mainly rely on masking or blurring, which can disrupt context and interaction understanding. We propose a generative, semantic-aware replacement approach that substitutes sensitive regions with visually coherent content in dynamic screen environments. We implement two variants with different degrees of semantic grounding: Semantic-Element-Preserving Replacement (SEPR) and Semantic-Element-Agnostic Replacement (SEAR), and compare them with masking and blurring. In a video-based user study that isolates latency effects, we evaluate detectability, confidence of detection, perceived identifiability, and user experience from both screen-sharer and screen-viewer perspectives. SEAR significantly reduced detectability relative to masking, and also lowered perceived identifiability compared to blurring and SEPR. Replacement methods significantly improved context understanding and visual naturalness over masking, and SEAR also outperformed blurring in visual naturalness; for context preservation, SEAR was rated higher than masking, while the SEPR–masking comparison was near the significance threshold. Qualitative responses further revealed role-dependent concerns and a social risk of misinterpretation, motivating design implications for tuning semantic fidelity and signaling obfuscation intent.
Authors
- Yuki Igarashi (ORCID: https://orcid.org/0000-0001-5025-8526)
- Kaori Ikematsu (ORCID: https://orcid.org/0000-0002-7017-6744)
- Rimi Makita (ORCID: https://orcid.org/0009-0003-5398-9350)
Institutions
- Ochanomizu University (JP)
- The University of Tokyo (JP)
Publication Details
- Journal
- Proceedings of the ACM on Human-Computer Interaction
- Published
- 2026-11-20
- DOI
- https://doi.org/10.1145/3859318
- Primary Topic
- Privacy, Security, and Data Protection
- Type
- article
- Field-Weighted Citation Impact
- 0.00