Minimax optimal differentially private synthetic data for smooth queries
Researchers have developed a new algorithm for generating differentially private synthetic data that offers improved utility for specific types of queries. The algorithm achieves a minimax error rate of O(n^{-min{1, k/d}}) for k-smooth queries, outperforming previous methods for these specific query classes. This work also establishes the first minimax lower bound for k-smooth query utility under $(\varepsilon, \delta)$-differential privacy. AI
IMPACT Enhances privacy-preserving data analysis techniques, potentially enabling broader use of sensitive datasets for research.