Glossary · Business Analytics

Differential Privacy

also: DP · ε-differential privacy · epsilon-DP · Laplace mechanism · noise injection

Definition

Differential privacy is a mathematical framework (Dwork 2006) that bounds the privacy loss any individual incurs from a query by adding calibrated random noise. The privacy budget ε quantifies the trade-off: smaller ε gives stronger guarantees and noisier outputs; production deployments typically run with ε between 1 and 10.

Differential privacy (DP) provides a formal guarantee that no individual record materially influences the output of any analysis. The Laplace and Gaussian mechanisms add noise calibrated to the query's sensitivity (the maximum change one record can produce). The privacy budget ε bounds cumulative privacy loss across queries; (ε, δ)-DP relaxes the guarantee with a small probability δ of unbounded loss. Production deployments include Apple's keyboard-suggestion telemetry, the US Census Bureau's 2020 disclosure-avoidance system, and Google's RAPPOR (later Prochlo). DP is appropriate for aggregated metrics over large populations; it is poorly suited to individual-row analytics where the noise overwhelms the signal.

Essays on this concept