Abstract
Differentially private algorithms for common metric aggregation tasks, suchas clustering or averaging, often have limited practicality due to theircomplexity or a large number of data points that is required for accurateresults. We propose a simple and practical tool $\mathsf{FriendlyCore}$ thattakes a set of points ${\cal D}$ from an unrestricted (pseudo) metric space asinput. When ${\cal D}$ has effective diameter $r$, $\mathsf{FriendlyCore}$returns a "stable" subset ${\cal D}_G\subseteq {\cal D}$ that includes allpoints, except possibly few outliers, and is {\em certified} to have diameter$r$. $\mathsf{FriendlyCore}$ can be used to preprocess the input beforeprivately aggregating it, potentially simplifying the aggregation or boostingits accuracy. Surprisingly, $\mathsf{FriendlyCore}$ is light-weight with nodependence on the dimension. We empirically demonstrate its advantages inboosting the accuracy of mean estimation, outperforming tailored methods.