Differentially Private Histograms for Range-Sum Queries: A Modular Approach.

We focus on the problem of differentially private histogram publication, for range-sum query answering. Specifically, we derive a histogram from a given dataset, such that (i) it satisfies $\epsilon$-differential privacy, and (ii) it achieves high utility for queries that request the sum of contiguous histogram bins. Existing schemes are distinguished into two categories: fast but oblivious to utility optimizations that exploit the data characteristics, and data-aware but slow. We are the first to address this problem with emphasis on both efficiency and utility. Towards this goal, we formulate a principled approach, which defines a small set of simple modules, based on which we can devise a variety of more complex schemes. We first express the state-of-the-art methods in terms of these modules, which allows us to identify the performance bottlenecks. Next, we design novel efficient and effective schemes based on non-trivial module combinations. We experimentally evaluate all mechanisms on three real datasets with diverse characteristics, and demonstrate the benefits of our proposals over previous work.