PerceptFence: Content-Mediation Architecture and Deterministic Coverage for Screen-Share AI Assistants
PerceptFence mediates screen-share AI assistant capture to redact secrets, beating Presidio on OCR-surviving PII in synthetic tests.
PerceptFence is a content-layer mediation architecture between screen and speech capture, memory, and model responses for live screen-share AI assistants. The artifact is a deterministic synthetic-fixture scaffold and omits live capture, category inference, authenticated re-consent, cross-session state, and an external model adapter. On 9,600 adversarial strings it neutralized 0.828 of digit-PII payloads versus 0.183 for Microsoft Presidio on shared seeds, while Presidio led outside that family. On 480 synthetic Chrome screens it neutralized 889 of 968 OCR-surviving secrets and PII values (0.918), versus 0.581 for Presidio and 0.179 for gitleaks, with 0.763 task-token retention on held-out types.
- Mediation sits between capture, memory, and model responses.
- Artifact omits live capture and an external model adapter.
- It neutralized 889 of 968 OCR-surviving secrets and PII values.
- Digit-PII beats Presidio; Presidio leads outside that family.
- Held-out screens scored 0.974 with 0.763 token retention.
From the original article
1 min read at arxiv.orgLive screen-share AI assistants observe raw screen and speech streams, but users have little runtime control over what an assistant may observe, retain, or disclose. Prompt-level privacy settings are insufficient because sensitive content enters through the capture stream.
By Asmita Negi · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.