Secure support vector machines with data perturbation
Xinning Li, Zhiping Zhou · 2018
In view of the increasing demand for privacy protection, traditional data mining process has to be optimized when the data owners are usually unwilling to release their original data for analysis. Aiming at the data classification in data mining, we have proposed CI-SVM (Condensed Information-Support Vector Machine) algorithm to achieve safe and efficient data classification. In this paper, the RCI-SVM (Random Linear Transformation with Condensed Information-Support Vector Machine) algorithm is proposed to use random linear transformation to convert the condensed information to another random vector space. The compressed information in CI-SVM are obtained by clustering the original data, although it is possible to ensure that the accurate original information will not be exposed, to some extent they may still carry some characteristics of the original datasets. Unlike most of the existing data perturbations, due to the early information enrichment processing, RCI-SVM will not preserve the dot product and Euclidean distance relationship between the original datasets and the transformed datasets, which means it's stronger than existing methods in security. Our experiment results on datasets show that the proposed RCI-SVM algorithm can performs well on classification efficiency and security.