PrivScope: On-Device Privacy for Hybrid AI Agents
PrivScope is an innovative system designed to tackle the issue of excessive information sharing in hybrid local-cloud AI agents. These agents enhance user queries with context from their ongoing state prior to transmitting subtasks to a cloud language model (CLM), which can lead to the unnecessary exposure of data. Acting as a reliable on-device payload manager, PrivScope regulates task-specific disclosures at the local-CLM interface without necessitating modifications on the cloud side. It guarantees that sensitive data is only sent to the cloud when essential for the assigned subtask, and in the most minimal form that maintains its usefulness. This system was detailed in a paper published on arXiv (2605.16630).
Key facts
- PrivScope is a trusted on-device payload governor.
- It enforces task-scoped disclosure at the local-CLM boundary.
- No cloud-side changes are required.
- Sensitive information is only sent when necessary for the subtask.
- Information is sent in the least revealing form preserving utility.
- Hybrid local-cloud agents enrich requests with context from persistent working state.
- Over-disclosure includes task-irrelevant context, carryover from prior workflows, and overly specific sensitive details.
- Existing solutions isolate workflows or apply general-purpose sanitization.
Entities
Institutions
- arXiv