Differential Privacy

Differential privacy adds carefully calibrated statistical noise to a computation so that its result is nearly identical whether or not any one person's record is included. That bound is expressed as a privacy budget (epsilon): the smaller the budget, the stronger the guarantee and the more noise applied.

It matters for AI because models and synthetic-data generators can otherwise memorize and leak individual records. Differential-privacy controls let a team tune the trade-off between statistical fidelity and confidentiality, so a dataset can be shared or a model trained without exposing the people in it.