What is differential privacy?
Differential privacy is a mathematical framework that bounds how much one person’s inclusion can affect a released result.
How it works and where it is used
Carefully chosen noise can be added to aggregate counts or model training. Parameters such as epsilon describe protection, but their meaning and privacy budget across queries must be stated. It is not the same as simply saying data was anonymized. Repeated releases can accumulate privacy loss.
Limits, safety and selection criteria
Stronger privacy can reduce accuracy for small groups. Review the threat model, implementation bugs and links to other datasets. Even a sound mathematical mechanism cannot compensate for poor collection and access policies. A differentially private release limits how much one person can affect the output. Many methods add calibrated random noise to an aggregate. A city might publish trip counts while making it harder to infer whether a particular person took a trip. The privacy budget controls a tradeoff between privacy and precision. Repeated queries can consume that budget, so a dashboard needs query accounting rather than an unlimited export button. State the threat model, budget and expected uncertainty when publishing results.