The war on statistical noise · ↗ www.commerce.gov

The US Department of Commerce recently announced new rules governing disclosure avoidance in statistical products from the Census Bureau and the Bureau of Economic Analysis. They are banning noise infusion (a differential privacy technique) in favour of coarsening, or suppression as a last resort:

a. Coarsening shall be the preferred category of Disclosure Avoidance methods for all statistical products. b. Suppression shall be permitted as a last resort, only to be used when coarsening is prohibited by law or would substantially defeat the accuracy or usability of a statistical product. c. Noise infusion shall not be used for any statistical product.

Noise infusion preserves granularity by distorting values in a mathematically controlled way. The new preferred solution is simpler: provide less detail through aggregation and rounding. You can make the case for either method, though differential privacy techniques at least make the trade-offs more measurable.

I find the politics of this a bit mysterious (is differential privacy woke now?). The technical case against noise infusion is obvious enough: its use in the 2020 Census was legitimately controversial, particularly for the smallest geographies. But this justifies caution, not a wholesale ban. A categorical prohibition sends a message, even if I’m not quite sure what it is.