Private Sampling: A Noiseless Approach for Generating Differentially Private Synthetic Data

In a world where artificial intelligence and data science become omnipresent, sharing is increasingly locking horns with data-privacy concerns. Differential privacy has emerged as rigorous framework for protecting individual in statistical database, while releasing useful information about the database. The standard way to implement differential inject sufficient amount of noise into data. However, addition other limitations privacy, this process adding will affect accuracy utility. Another approach enable based on concept synthetic goal create an as-realistic-as-possible dataset, one that not only maintains nuances original data, but does so without risk exposing sensitive information. combination been suggested best-of-both-worlds solutions. work, we propose first noisefree method construct differentially private data; do through mechanism called sampling. Using Boolean cube benchmark model, derive explicit bounds constructed key mathematical tools are hypercontractivity, duality, empirical processes. A core ingredient our sampling marginal correction method, which remarkable property importance reweighting can be utilized exactly match marginals sample population.