SPLICE achieves dimension-independent privacy guarantees for synthetic data by matching polynomial moments along random directions, solving a major limitation of existing DP methods that either focus on marginals or suffer exponential dimension dependence.
This paper introduces SPLICE, a differentially private synthetic data algorithm that avoids the curse of dimensionality by using sliced Chebyshev polynomial moments. Unlike prior methods that only preserve low-dimensional patterns, SPLICE provides global distributional guarantees with error that doesn't grow with data dimension, while achieving state-of-the-art practical performance.