Differentially Private Principal Component Analysis Over Horizontally Partitioned Data
Sen Wang, J. Morris Chang · 2018
Principal Component Analysis (PCA) is widely adopted in various data mining and machine learning applications, it computes a low dimension subspace that captures the most variances of the underlying data. The area of distributed computing provides a promising domain for PCA, where it has been studied in many fields. In big data era, large volume and high dimensional data are generated at all times. For instance, mobile devices become the important producer and carrier for personal information, which can provide a considerable social utility. However, the current distributed PCA protocol cannot provide the efficiency and scalability with respect to such large amounts of data. Furthermore, the privacy issue arises when data contains sensitive information. The data owner would not prefer to sharing the data in cleartext, and the inference from PCA should also be prevented. Motivated to resolve these challenges, in this paper, we design and implement a highly efficient and largely scalable privacy preserving distributed PCA protocol, in which the (ε, δ)-Differential Privacy is guaranteed. In the experiments, we evaluate the protocol in terms of efficiency and utility, and shows that it maintains a high data utility while preserving the privacy.