Differentially Private Kernel Support Vector Machines Based on the Exponential and Laplace Hybrid Mechanism
Algorithm 1
DPKSVMEL
Input: Q: symmetric kernel matrix; ɛ: privacy budget; LLs: lower limit of the Similarity; Nns: the number of non-SVs in a group; k: the number of non-SVs selected in the exponential mechanism;
Output: SVp: private SV;
Begin
(1)
obtain a non-private classification model including dual vector α and the SVs by training a kernel SVM;
(2)
get the Similarity matrix from the subset of Q in which the Similarity value was no less than LLs;
(3)
divide every non-SV into one group according to the maximal value of its similarity with every SV;
(4)
for i in every group
(5)
if Nns > k then
(6)
compute the probability Prns for every non-SVs with its Similarity value;
(7)
randomly select the most similar k non-SVs with probability Prns by the exponential mechanism;