Designing for User Privacy: Integrating Differential Privacy into Ad Measurement...
summary
A practical system integrates distributed differential privacy with MPC/TEE-based matching to protect both intermediate and final ad-attribution outputs, adding a common dummy dataset (shuffled and appended) to hide intersection sizes. Without output DP the measurement can leak nearly 10% membership; prior intersection attacks leak ~0.4% and an improved attack reaches ~2.5%. A matrix-mechanism optimization minimizes total noise over streaming reports and a tight DP composition cuts dummy-size and noise, yielding better accuracy on public ad-campaign datasets while increasing communication and computation costs.
AI-written summary · about
chapters (6)
- 0:00Privacy in ad measurement
- 5:45Privacy-preserving ad landscape
- 7:34Privacy-enhancing technologies
- 9:02Integrating differential privacy
- 11:31Output privacy deep dive
- 18:55Key takeaways and future
related talks
19mPractical Considerations for Differential Privacy
20mShare data without revealing personal information
01h 05mMarkets and Incentives for Transacting on (Personal) Information, Part III
01h 10mDifferential Privacy: Fundamentals to Forefront II
01h 00mPANEL: Statistical Theory, Privacy and Data Analysis
21m"SECRETS ARE LIES, SHARING IS CARING, PRIVACY IS THEFT." - A Dive into Privacy Preserving
30mAccuracy First: Selecting a Differential Privacy Level for Accuracy-Constrained...
26mAaron Roth, Associate Professor, University of Pennsylvania, NYC 2017