Marginal-based Methods for Differentially Private Synthetic Data - Differential Privacy for ML Series

Explore marginal-based methods for generating differentially private synthetic data in this 43-minute Google TechTalk presented by Ryan McKenna. Delve into the NIST DP Synthetic Data Competition setup and understand the importance of marginals in data privacy. Learn about various mechanisms including independent baseline, MST selection algorithm, and comparisons between Bayesian Network and Markov Random Field approaches. Discover interesting empirical findings and considerations for selection, as well as budget-aware, workload-aware, data-aware, and efficiency-aware mechanisms. Gain insights into qualitative comparisons of prior work and explore summary and open problems in the field of differential privacy for machine learning.

Intro
NIST DP Synthetic Data Competition
Competition Setup
Marginal-based mechanisms
Why Marginals?
Independent Baseline
MST Selection Algorithm
Bayesian Network vs. Markov Random Field
Select the Workload?
Interesting Empirical Finding
Considerations for Selection
Budget-Aware Mechanism
Workload-Aware Mechanism
Data-Aware Mechanism
Efficiency-Aware Mechanism
Qualitative Comparison of Prior Work
Summary & Open Problems