Implementing Differential Privacy for the 2020 US Census

Explore the implementation of differential privacy for the 2020 U.S. Census in this USENIX Enigma Conference talk. Delve into the scientific, technical, and organizational challenges faced by the U.S. Census Bureau in bringing cutting-edge privacy technology from theory to practice. Learn about the development process, including assembling a team, creating a reference implementation, transitioning to Amazon Web Services, and redesigning the implementation framework. Discover how the Bureau conducted experiments, performed end-to-end testing, and re-released data from the 2010 Census using the new system. Gain insights into the evolution of differential privacy since its invention in 2006, the technical hurdles overcome, and the ongoing efforts to meet user expectations. Understand the complexities of deploying differential privacy in a large-scale statistical operation and the implications for future privacy-preserving data releases.

Intro
Differential Privacy was invented in 2006 (14 years ago) Modem Publik Key Cryptography was invented between 1976 and 1978
The Census Bureau deployed the first DP implementation in the world for OnTheMap (2008)
We solved many technical challenges deploying DP for the 2020 Census!
We still haven't met user expectations.
The Decennial Census!
By 2017, we thought we had a good understanding of how DP would fit.
Different groups at USCB had different assumptions about the data flow.
The DP system had to be developed with real (Title 13-protected) data. Most systems for the 2020 Census were developed using simulated data
Census Bureau leadership was 100% behind the move to Differential Privacy
Team and Reference Implementation
CPU load of 21 AWS Instances during an execution of the TopDown algorithm
We re-published the 2010 data using the 2020 TopDown algorithm.
At the root of our acceptance problem: Non-negativity requirements create systematic bias and increase the size of outlier errors
In summary: The 2020 Census DP Timeline