Noisette
Unified framework for certifying differential privacy mechanisms across discrete and continuous distributions
Proposed Noisette, a unified framework for certifying DP noise sampling across discrete and continuous mechanisms, supporting arbitrary additive noise including Gaussian, Laplace, and Skellam.
- Designed a certifiable lookup-table protocol for arbitrary distributions; achieves up to 30× runtime improvement and 36× communication reduction over prior SOTA for discrete Gaussian sampling.
- Provided the first better-than-naive certifiable protocol for continuous Gaussian sampling, achieving ~64× speedup and ~15× lower communication vs. SOTA.
Authors: Qi Pang, Radhika Garg, Ziling Liu, Hanshen Xiao, Virginia Smith, Wenting Zheng, Xiao Wang
Status: In submission
Paper: ePrint 2026/074