Differential privacy
By Gareth HoyleLast updated How we research
Differential privacy is a mathematical guarantee that limits how much any result reveals about whether a single person's data was included.
Differential privacy is a mathematical framework for measuring and limiting how much an individual's privacy is exposed when their data is included in a dataset or analysis. A process is differentially private if its output looks almost the same whether or not any one person's record is included, with the gap controlled by a privacy parameter usually written as epsilon.
What guidance does NIST give?
NIST published SP 800-226, Guidelines for Evaluating Differential Privacy Guarantees, in March 2025 to help practitioners understand and evaluate software that claims to implement it.
How is differential privacy used with training data?
Differential privacy is used to train models with noise added during training, to release statistics about a dataset, and to generate synthetic data with a measurable privacy bound. It reduces the risk that a model memorizes and repeats individual records.
It is not a substitute for de-identification of the raw data a seller delivers. A buyer that receives raw records gets no protection from differential privacy applied later, so sellers still need to remove identifiers before delivery.