A Statistical Viewpoint on Privacy - From the 2020 United States Decennial Census to Hypothesis Testing to Blackwell's Theorem

Explore the statistical foundations of privacy-preserving data analysis through a comprehensive examination of differential privacy and its applications to the 2020 United States Census in this 47-minute lecture. Discover how differential privacy can be understood as a purely statistical concept by examining its formal motivation through hypothesis testing and Blackwell's informativeness theorem. Learn about the development of f-differential privacy framework and its enhanced privacy analyses across various applications. Analyze the Census Bureau's adoption of differential privacy in their disclosure avoidance system and examine whether stronger privacy guarantees could be achieved beyond their published standards. Investigate findings that demonstrate the 2020 Census provides significantly stronger privacy protections than officially reported across eight geographical levels, revealing that unnecessarily high noise levels were introduced. Understand how noise variances could be reduced by 15.08% to 24.82% while maintaining equivalent privacy protection, thereby improving the accuracy of privatized census statistics without compromising privacy guarantees.