User-Level Differentially Private Mean Estimation for Real-World Datasets

In this work, we provide rigorous theoretical justifications for the performance trends of well-known clipping-based algorithms on real-world ITMS and i.i.d. synthetic datasets. An important contribution of this work is the formalization and explicit computation of the “worst-case estimation error” incurred by a canonical user-level differentially private algorithm for mean estimation, which clips the number of contributions of users in an attempt to increase the accuracy of reconstruction.