Correlated Differential Privacy of Multiparty Data Release in Machine Learning

Differential privacy (DP) is widely employed for the private data release in the single-party scenario. Data utility could be degraded with noise generated by ubiquitous data correlation, and it is often addressed by sensitivity reduction with correlation analysis. However, increasing multiparty data release applications present new challenges for existing methods. In this paper, we propose a novel correlated differential privacy of the multiparty data release (MP-CRDP). It effectively reduces the merged dataset's dimensionality and correlated sensitivity in two steps to optimize the utility. We also propose a multiparty correlation analysis technique. Based on the prior knowledge of multiparty data, a more reasonable and rigorous standard is designed to measure the correlated degree, reducing correlated sensitivity, and thus improve the data utility. Moreover, by adding noise to the weights of machine learning algorithms and query noise to the release data, MP-CRDP provides the release technology for both low-noise private data and private machine learning algorithms. Comprehensive experiments demonstrate the effectiveness and practicability of the proposed method on the utilized Adult and Breast Cancer datasets.