Differential privacy in the 2020 US census: what will it do? Quantifying the accuracy/privacy tradeoff

Area of Science:

• Data privacy and security
• Statistical methodology
• Demographic research

Background:

• The 2020 US Census utilizes the novel TopDown algorithm for disclosure avoidance.
• The TopDown algorithm was tested in the 2018 end-to-end census test.
• Census Bureau has publicly released the code and data from the 2018 test.

Purpose of the Study:

• To analyze the error introduced by the TopDown disclosure avoidance system.
• To empirically measure privacy loss in the TopDown approach.
• To compare TopDown's privacy and error with simple random sampling.

Main Methods:

• Utilized publicly available code and data from the 2018 end-to-end census test.
• Applied the TopDown algorithm to 1940 US census data.
• Developed an empirical measure of privacy loss for comparison.

Main Results:

• TopDown's empirical privacy loss is significantly lower than its theoretical guarantee.
• TopDown with a privacy budget of 1.0 showed error and privacy loss similar to a 50% simple random sample.
• TopDown with a privacy budget of 4.0 showed error and privacy loss similar to a 90% simple random sample.

Conclusions:

• The study contributes to the ongoing discussion on balancing privacy and accuracy in census data.
• Further research and discussion are needed to optimize data collection methods.
• The TopDown algorithm shows promise for effective disclosure avoidance with manageable error.