Differential privacy is a mathematical privacy standard that guarantees the presence or absence of any individual's record in a dataset cannot be inferred from analysis results or released data. It adds controlled noise to outputs so that the influence of any single record stays within a bound, expressed as a privacy budget (epsilon). The concept was introduced by Cynthia Dwork and colleagues in 2006.
Unlike de-identification that merely removes identifiers such as names, it offers a mathematically guaranteed defense against attacks that re-identify individuals by cross-referencing other data. In AI, common applications include applying differential privacy during training (DP-SGD) and generating differentially private synthetic data, which preserves the statistical properties of the original while severing links to individual records, for use in model training. Apple, Google and the U.S. Census Bureau have applied the technique to data collection and statistical releases.