Privid - Practical, Privacy-Preserving Video Analytics Queries

Explore a conference talk on Privid, a practical system for privacy-preserving video analytics queries. Delve into the challenges of balancing utility and privacy in public surveillance camera footage analysis. Learn about a new notion of differential privacy called (ρ,K,ε)-event-duration privacy, which protects private information visible for less than a specific duration. Discover how Privid enforces duration-based privacy while working with analyst-provided deep neural networks. Examine the system's performance across various videos and queries, comparing its error rates to non-private systems. Gain insights into the query interface, threat model, and key observations made during the research. Understand the potential applications and privacy implications of video analytics in public spaces.

Introduction
Surveillance cameras
Threat model
Differential privacy
The challenge of differential privacy
Observations
Takeaways
Query Interface