Privid: Practical, Privacy-Preserving Video Analytics Queries
Talkat 19th USENIX Symposium on Networked Systems Design and Implementation (NSDI '22) · 2022-04-04..04-06·
summary
Privid defines ρ,K,ε-event-duration privacy and enforces it by splitting video into chunks, running analyst-provided black-box neural nets per chunk, then split-process-aggregate with a constrained execution environment and bounded schema so each person can only affect rows while visible; anything visible for less than ρ seconds (cars ~3–5s, people ~60s on average) gets protected. The system adds DP noise proportional to the computed sensitivity, supports cross-camera joins (e.g., license-plate intersections), and raises error only 1–21%.
AI-written summary · about
chapters (8)
- 0:00Introduction
- 0:28Surveillance cameras
- 1:40Threat model
- 5:05Differential privacy
- 7:33The challenge of differential privacy
- 8:26Observations
- 9:18Takeaways
- 14:54Query Interface
related talks
16mPrvTel: Lightweight Models for Private and Accurate Telemetry Data Retention
14mToward Provably Private Insights into AI Use
32mPrivacy Protection. Privacy by Design: Strengthening Compliance
45mMaking Privacy Technology Accessible: Benchmarks and Platforms
22mEnhancing Privacy through an Interactive On-demand Incremental Information
20mShare data without revealing personal information
26mAaron Roth, Associate Professor, University of Pennsylvania, NYC 2017
19mPractical Considerations for Differential Privacy