Data-dependent differentially private publication of horizontally partitioned data
Ning Wang · 2022
Probabilistic graphical model (PGM) is a mainstream model for data publishing. Due to its structural characteristics, the higher the accuracy of important nodes in PGM, the higher the utility of published data. This paper proposes a datadependent differentially private publishing method for horizontally partitioned data, including a multi-party data publishing framework which can combine the existing data-dependent parameter learning method and a data-dependent structure learning method in the horizontally partitioned data setting. So that the whole learning process of PGM is datadependent. Thus, different privacy budgets can be allocated to different nodes when learning the PGM for different datasets. The experimental results show that the data published by our scheme has high utility.