A Unified Model for Privacy-Preserving Support Vector Machines on Horizontally and Vertically Partitioned Data

摘要

We propose a novel unified model for Privacy-Preserving Support Vector Machines (PPSVM for short) classifier on horizontally and vertically partitioned data. We prove the feasibility of the model. Besides we give out the algorithms for horizontally partitioned data and vertically partitioned data, respectively. The columns of data matrix A represent input features and the rows represent the individual data which is called a training/testing point in SVM. For horizontally partitioned data, the data matrix A whose rows including all input features are divided into groups belonging to different entities. While for vertically partitioned data, the data matrix A`s columns are divided into groups belonging to different entities. Each entity is unwilling to share its group of data or leak the data for various reasons. The proposed SVM classifiers are public but do not reveal any private data. And when we calculate the classifier at last, we do not need to recover the original data. Besides, it has comparable accuracy with that of an ordinary SVM classifier that uses the centralized data set directly. Experiments show that our approach is effective.