A mathematical definition of privacy under which the output of an analysis changes only marginally when any single individual's record is added to or removed from the input dataset, quantified by a privacy budget epsilon.
In practiceDifferential privacy provides a worst-case guarantee against re-identification that holds regardless of the attacker's auxiliary knowledge, by injecting calibrated noise into queries, gradients, or released statistics. Smaller epsilon values give stronger privacy at the cost of accuracy. It is deployed in census releases, telemetry collection, and privacy-preserving machine learning, and is increasingly cited by regulators as a recognised technique for reconciling model utility with data protection obligations. Composition rules govern how repeated queries consume the budget.
A health insurer trains a hospitalisation risk model using differentially private stochastic gradient descent with an epsilon of three across the training run, allowing it to publish aggregate model statistics with a defensible bound on what any individual policyholder's record contributed.
This definition is maintained by Moweb partners and used in live client engagements. For how Differential privacy applies to your estate, or to challenge a working definition, speak to a partner.