Differential Privacy as a Causal Property

Differential privacy can be characterized as a causal property: without assuming that data points are independent of each other, it can still be formulated as limiting the influence of a single data point. From a correlation perspective, dependence among data points makes it impossible to simply formulate differential privacy as conditioning on a single changing data point; the causal perspective provides a characterization that does not rely on the independence assumption. The authors argue that this can clarify previous disagreements about the consequences of differential privacy and make it possible to apply results about causality from statistics, experimental design and science to differential privacy research.