Differential Private Multiple Classification Algorithm for SVM
Han Wang, Shuyu Li · 2018
Data mining algorithm has privacy leakage problem. To solve this problem, a privacy preserving support vector machine (SVM) algorithm under differential privacy for multiple classification is proposed in the paper. The algorithm disturbs the kernel function by three different ways including direct Laplace noise injection, Taylor formula replacement, and combination of previous two methods. The whole classification model is disturbed indirectly by the value of the normal vector obtained from disturbance. It is expected to protect the small sample data and not to interfere with the classification effect of the model to the whole data set. During experiment phase, the algorithm is verified by linear kernel, polynomial kernel and Gaussian kernel. The experimental results show that the algorithm can achieve data utility of classification result while satisfying privacy protection requirements.