Reproduce “VertiMRF: Differentially Private Vertical Federated Data Synthesis” on OpenResearch

About this paper

Fangyuan Zhao, Zitao Li, Xuebin Ren, Bolin Ding, Shusen Yang, Yaliang Li

The VertiMRF framework, developed by researchers at Xi'an Jiaotong University and Alibaba Group, introduces a differentially private data synthesis method for vertical federated learning that leverages local Markov Random Fields and Flajolet-Martin (FM) sketches for cross-party correlation estimation. It consistently outperforms existing baselines like VertiGAN, achieving lower Total Variation Distance and misclassification rates on real-world datasets, while managing communication and computation costs.

Key insights

• The Flajolet-Martin (FM) sketch-based approach provides more accurate cross-party marginal estimations than traditional Frequency Oracle methods in VFL DP synthesis.
• Leveraging local MRFs to capture intra-party correlations significantly improves utility and reduces result variance in the overall synthetic data generation.
• Dimension reduction through binning, coupled with histogram recovery, and consistency enforcement are crucial for maintaining utility in high-dimensional datasets and with varying noise sources.

Summary by alphaXiv.