Differential Privacy at the US Census

Concept & History

Differential privacy, a concept introduced by Cynthia Dwork in 2005, aims to protect individual data contributions by adding noise to the data, ensuring privacy without compromising statistical accuracy. Simson Garfinkel explains that this technique allows for a balance between privacy loss and data utility, controlled by a parameter called epsilon 1. The Census Bureau, pioneering this approach since 2008, has faced challenges in transitioning from traditional methods to differential privacy, requiring significant technological advancements and policy changes 2.

   

Implementation

Implementing differential privacy in the census involves overcoming technical and logistical hurdles, particularly in post-processing data to maintain consistency. Garfinkel highlights the computational demands, requiring a cluster with over 1,000 cores running for more than 36 hours to process the 2020 census data 2. Despite these challenges, the Census Bureau has made strides in transparency by publishing source code and engaging with external consultants to refine their methods 3.

   

Data Impact

Differential privacy impacts data accuracy, with noisy measurements offering a new level of transparency and complexity. Garfinkel notes that while the final statistics may appear less accurate than previous methods, they provide a more secure privacy framework 4. The introduction of differential privacy requires data users to adapt to new data formats and understand the underlying mechanisms to fully leverage the benefits 5.