Definition
Differential privacy is a mathematical guarantee about a randomized computation. Informally, the output should look nearly the same whether one person's data is included or left out. Cynthia Dwork, Frank McSherry, Kobbi Nissim, and Adam Smith introduced the formal framework in 2006.
The guarantee is stated over neighboring datasets, which differ in the protected contribution of one person or unit. Epsilon describes the allowed privacy loss; approximate differential privacy also uses delta for a small probability of a larger deviation. Smaller values usually provide stronger privacy but require more noise or some other loss of utility.
Privacy spending composes. Releasing several answers about the same people can reveal more than releasing one, so a serious claim states the unit of protection, epsilon and delta, the neighboring-dataset definition, clipping or sensitivity rules, and the total budget across releases.
Distinguish it from nearby terms
De-identification removes or masks identifiers. Encryption protects data in storage or transit. Differential privacy limits what can be inferred from a released computation under a quantified privacy budget. It does not promise zero leakage, and adding unspecified noise is not enough to claim the guarantee.
Check your understanding
A dashboard says it is "privacy preserving" because it rounds small counts. That is not yet a differential privacy claim. The team must define neighboring datasets, calibrate the mechanism, and account for repeated queries before the term applies.