Overview
Explore a programming framework for differential privacy that focuses on accuracy concentration bounds in this IEEE conference talk. Delve into DPella, a novel approach that enables data analysts to reason about privacy, accuracy, and their trade-offs. Learn how the framework leverages taint analysis to infer statistical independence of noise quantities, resulting in tighter accuracy estimations. Discover the implementation of classical queries and how to calibrate privacy to meet accuracy requirements. Gain insights into differentially private queries, API components, compositional error bounds, and the comparison between Union and Chernoff bounds. Understand the limitations and potential extensions of this approach, equipping yourself with valuable knowledge for developing privacy-preserving data analyses.
Syllabus
Intro
Motivation
Differentially private queries
@ Differential privacy
@ API components (1)
Accuracy of the analyses
Compositional error bounds
Union vs Chernoff bound
Limitations and Extensions
Conclusions
Taught by
IEEE Symposium on Security and Privacy