Support Vector Machine Algorithm With Differential Privacy Protection

Thanks to the development of information technology,there is no place for everyone to hide from the exquisite and rich data.How to carry out data research safely,under those multi-dimensional and fine-grained data with hidden sensitive information,is a problem worthy of discussion.Although the traditional privacy protection model can meet the needs of privacy protection,it has two disadvantages: one is that it can't quantify the balance between effectiveness and privacy;the other is that it's limited by the background knowledge reserve of the attacker.In view of this,our paper intro-duces a differential privacy protection technology with privacy parameters and presets the maximum background knowledge of the attacker to solve the above two problems.In this paper,the optimal classification hyperplane normal vector of support vector machine classification model based on hinge loss is used as the entry point.The privacy protection of classification task is realized in two steps.First,obtaining the iterative ex-pression of the normal vector (?) by locally smoothing the objective function of the SVM classification model.Second,basing on the sequence combination property of differen-tial privacy protection and the Gaussian implementation mechanism,defining the noise distribution over the iteration process,so as to obtain the parameter iteration under dif-ferential privacy protection.Subsequently,our paper proves that the algorithm is(?,?)differential private and extends it to pinball loss-based support vector machine classi-fication model.Experimental results show that the algorithm performs better when the sample size is sufficient.At the same time,compared with the non-privacy protection algorithm,the proposed algorithm still has good classification ability when it has a cer-tain privacy budget.Compared with the traditional privacy protection algorithms,the results of the proposed algorithm are more stable.