Differential privacy is a cornerstone of modern-day
#statistics.
It allows the generation of datasets that are statistically relevant for a researcher, but that leak as little as possible about the individuals in the dataset.
It works on several principles:
Suppression, usually by removing outliers from the dataset that could lead to easily identifiable targets.
Coarsening, e.g. replacing a city with a region or a country, or a date of birth with an age range.
Sampling, usually through the removal of random records.
Swapping, where some attributes from different records may be randomly swapped.
Noise addition, by adding or subtracting random numbers in a given noise range to the actual values.
From 1990 to 2010, the US Census Bureau primarily relied on swapping for the decennial census. Then, they realized that this technique was actually very unsafe, and that it was pretty easy to reconstruct individual records using the published statistics.
They eventually adopted the full differential privacy framework in 2020.
But some people apparently got annoyed with that. And the main culprit was noise addition.
Adding a normal random distribution to the data made the numbers noisy, and some people who rely on the US Census expecting its numbers to be fully accurate were very annoyed.
Who are these people?
Well, one category is that of demographers and social scientists. Still, their mathematical models can be adapted to take into account that they’re dealing with noisy data - especially if the noise gain of distribution is published upfront.
But the most annoyed ones are thos who used the Census data to reconstruct actual records of individuals.
Yes, you read it well. There are people who use public datasets that are supposed to be coarse, and leverage probabilistic models to reconstruct information about individuals.
One big use case in America is, of course,
#gerrymandering.
Those nonsense spiky borders can be much more accurate when the Census data allows you exactly to identify who lives where and what’s their ethnicity, so you can dilute the voting powers of some ethnic groups much more efficiently.
Apparently those groups have lobbied the Federal government so much that they got it to release an order that states the following:
Any use of noise infusion is inconsistent with the Department’s policies
This is a serious attempt of a political body to interfere through legislation into technical decisions that should be uniquely delegated to technical specialists - and the academic consensus is currently that of a full differential privacy framework.
Noise infusion is the most effective pillar in the differential privacy framework to prevent identification of individuals in public datasets.
Weakening it is likely to make everyone’s
#privacy worse, and benefit only very few and their interests.
#USPolhttps://desfontain.es/blog/banning-noise.html