Covariance’s Loss is Privacy’s Gain: Computationally Efficient, Private and Accurate Synthetic Data

Abstract The protection of private information is vital importance in data-driven research, business and government. conflict between privacy utility has triggered intensive research the computer science statistics communities, who have developed a variety methods for privacy-preserving data release. Among main concepts that emerged are anonymity differential privacy. Today, another solution gaining traction, synthetic data. However, road to paved with NP-hard problems. In this paper, we focus on challenge develop generation method computationally efficient, comes provable guarantees rigorously quantifies utility. We solve relaxed version problem by studying fundamental, but first glance completely unrelated, probability concerning concept covariance loss. Namely, find nearly optimal constructive answer question how much lost when take conditional expectation. Surprisingly, excursion into theoretical produces mathematical techniques allow us derive constructive, approximately solutions difficult applied problems microaggregation,