Distributed Data Synthesis under Local Differential Privacy
Xiaoguang Li, Haonan Yan, Qingwen Li, Hui Li, Fenghua Li · 2024
Local Differential Privacy (LDP) has emerged as a cornerstone for safeguarding privacy in data analysis, enabling robust statistical inference without the need for a trusted intermediary. However, current LDP methods often lack scalability and applicability beyond specific tasks. In response, we propose a novel approach for locally differentially private data synthesis. Our method extracts representative marginals from user data to construct synthetic datasets that maintain statistical fidelity while adhering to differential privacy constraints. A primary challenge lies in selecting meaningful marginals and minimizing noise in LDP mechanisms tailored to domain size of each marginal. To tackle this, we employ a sliding overlapping window technique to effectively capture low-degree marginals that encapsulate critical dataset attributes. Furthermore, we adaptively utilize categorical frequency oracles (CFOs) to estimate marginal frequencies based on their respective domain sizes, thereby reducing noise levels and maintain data utility. Empirical evaluations on real-world datasets, including Twitter and Adult datasets, underscore the efficacy of our approach in preserving data utility across a spectrum of fundamental statistical tasks.