Differential Privacy Work Set Selection Method Based on Exponential Mechanism for Linear Support Vector Machines

In the era of big data, data privacy protection is of utmost importance. When Support Vector Machines (SVM) process sensitive data, there is a risk of privacy leakage. This paper proposes a Differential - Privacy Work Set Selection method based on the Exponential Mechanism (DPWSS), aiming to protect the privacy information in linear SVM classification models. First, this paper introduces the crucial role of work - set selection in SVM training and the limitations of existing methods. Subsequently, the DPWSS is elaborated in detail. Based on the WSS3 algorithm, this method designs a scoring function based on the gradient information of sample pairs. By combining with the exponential mechanism, it introduces randomness into the work - set selection process. It does not require the differentiability of the objective function nor complex sensitivity analysis. Then, the description of the DPWSS algorithm is given, and it is integrated into the SMO algorithm to construct a complete privacy - protected SVM training framework. Theoretical analysis proves that the DPWSS algorithm meets the requirements of differential privacy. Simulation experiments are conducted on the A1a dataset using the LIBSVM software package and the Python environment. The results show that under different settings of privacy budgets and constant factors, the classification performance of the models trained by DPWSS is minimally affected by randomness and is similar to that of the models trained by the WSS2 algorithm. Moreover, the work sets are stable, and the execution efficiency is relatively high. When the privacy budget is fixed, the number of iterations significantly decreases as the constant factor increases. In conclusion, DPWSS achieves a balance between privacy protection and classification performance, providing a new solution for privacy protection in the field of machine learning. However, there is still room for improvement in aspects such as optimizing the scoring function, perfecting the randomness - introduction mechanism, and extending to nonlinear SVMs.