"Differential Privacy the Correct Choice for the 2020 US Census"
The US Census Bureau has always sought to strike a balance between the accuracy and privacy of its decennial census data. The accuracy of this data is especially important in high-impact use cases such as budget allocation and redistricting. Census data privacy, on the other hand, is required by law and is critical for protecting vulnerable populations and guaranteeing a high response rate. The Census Bureau has moved to a new de-identification approach called Differential Privacy (DP) because of major privacy concerns regarding its previous de-identification method, swapping. DP is a mathematical concept that injects small random changes, known as noise, into data to keep people's personal information private. There have been concerns that this noise may artificially deflate minority group populations, resulting in funding cuts. Therefore, a team of researchers at Columbia Engineering's Department of Computer Science studied these claims, questioning whether this risk is unique to DP. They compared DP algorithms to swapping algorithms in both theoretical and empirical studies. Their findings support the Census Bureau's decision to use DP as a de-identification strategy for the 2020 Census and show that swapping results in low accuracy for minority groups. In addition, swapping was found to place a disproportionate privacy burden on minority groups, whereas DP gives a stronger guarantee. The researchers found that the additional inaccuracy introduced by swapping is more damaging than DP. They demonstrated that when swapping is implemented with enough privacy, its accuracy is no better than, and often worse than, DP. For more diverse counties, and especially more so for minorities, swapped data is more inaccurate; however, this is not the case for DP. This article continues to discuss findings from the Columbia Engineering researchers' study on the impact of DP and swapping on privacy preservation and minority representation in the US census.
Newswise reports "Differential Privacy the Correct Choice for the 2020 US Census"