A mathematical guarantee that a statistical result barely changes whether or not any one person is in the data, obtained through calibrated noise.
Differential privacy is a property of a computation rather than a technique for anonymizing data. It guarantees that the distribution of results produced is almost identical whether or not a given individual appears in the dataset, the maximum divergence being bounded by a publicly declared parameter. It follows that an observer, even holding all other information, cannot conclude that the person is present, a far stronger guarantee than removing direct identifiers, which has been defeated many times by linkage. Its cost is explicit and that is its virtue: the noise added degrades the accuracy of results, and the parameter fixes a trade-off made public rather than buried in a procedure. A budget is moreover consumed at every query, which forces uses to be counted. In machine learning the property is obtained during training and then bounds what a membership inference attack can learn, which makes it the only demonstrable rather than merely observed protection.
The United States Census Bureau applied differential privacy to the publication of the 2020 decennial census results, the first implementation at that scale, a decision challenged in court by states on the ground of lost precision in local data.
differential privacy, confidentialité différentielle, budget de confidentialité, epsilon