Research On Support Vector Machine Classification Algorithms Based On Differential Privacy Protection

Support Vector Machine(SVM)is one of the most widely used and efficient methods in the field of data mining classification.However,using data mining technology such as S VM to mine and analyze existing data sets will reveal individual privacy information.Differential Privacy(DP),as a rigorous and provable privacy protection technology,has been widely studied and applied to data publishing and data mining.Aiming at the privacy leakage problem of support vector machine classification algorithm,this dissertation studies the support vector machine classification algorithm based on differential privacy protection.The main research contents are as follows:1)In view of the low prediction accuracy and poor versatility of the existing SVM classifiers with privacy protection,a Differential Privacy Support Vector Machine algorithm based on Dual Variable Perturbation(DPSVMDVP)is proposed.The algorithm in this dissertation is different from the way of adding random perturbations in the existing literature,in order to improve the prediction accuracy of the support vector machine classification algorithm based on differential privacy.The algorithm solves the dual problem of the support vector machine.According to the difference Ei of the difference E between the estimated value and the real value corresponding to each support vector,the dual variable value αi corresponding to each support vector finally released is added with different levels of pull.Plass noise.The global sensitivity of DPSVMDVP algorithm when adding noise is given in the form of lemma,and the differential privacy of DPSVMDVP algorithm is proved.Finally,the experimental verification is carried out on the real data set.The result shows that DPSVMDVP algorithm is under reasonable privacy budget.The classification prediction accuracy is higher.2)Aiming at the privacy protection requirements of multi-class support vector machine classification algorithm,based on the multi-class SVM of one-versus-one method,a Differential Privacy Multi-Class Support Vector Machine(DPMCSVM)is proposed.First,the study defines a two-level weighted privacy budget allocation scheme and assigns a corresponding proportion of privacy budget to each sub-SVM classifier.The strategy of adding Laplace noise using the DPSVMDVP algorithm is used to add noise to the dual variables of each sub-SVM classifier,and a support vector machine multi-classifier satisfying differential privacy is obtained.Combining the property of differential privacy with the analysis of DPMCSVM algorithm’s privacy protection process,it is proved that DPMCSVM algorithm satisfies ε-differential privacy.Experiments on real data sets show that DPMCSVM algorithm can classify and predict better under reasonable privacy budget.