Privacy Preserving SVM with Different Kernel Functions for Multi-Classification Datasets
Z. Li, S. Li
Aiming at the mining problem of privacy preserving data, a SVM algorithm under differential privacy is presented for multi-classification datasets in the paper. The main idea of the proposed algorithm is to add Laplace noise for decision function, thus the privacy can be protected when the computation value of kernel function is changed and the normal vector is disturbed. Three different kernel functions including the linear kernel, the polynomial kernel and the Gaussian kernel respectively, are selected for classification comparison. Experiment results show that
three kernel functions can achieve better classification accuracy rate to a certain degree. From the view of the computation time, the liner kernel is the fastest while that of Gaussian kernel is
the slowest. From the perspective of classification accuracy rate after noise addition, the polynomial kernel is the most accurate.