DPWSS: differentially private working set selection for training support vector machines.

Area of Science:

• Machine Learning
• Data Privacy
• Computational Science

Background:

• Support Vector Machines (SVMs) are effective for classification but risk privacy when training data is sensitive.
• Traditional SVM training methods, particularly working set selection, can expose personal information.
• Existing privacy-preserving SVMs face challenges with complex sensitivity analysis and high-dimensional data noise.

Purpose of the Study:

• To propose a novel differentially private working set selection algorithm (DPWSS) for SVMs.
• To address privacy concerns in SVM training without compromising classification performance.
• To enhance the efficiency of privacy-preserving SVM training.

Main Methods:

• Developed a differentially private working set selection algorithm (DPWSS).
• Utilized the exponential mechanism for private selection of working sets.
• Provided theoretical proof of differential privacy for the DPWSS algorithm.

Main Results:

• DPWSS achieves classification performance comparable to non-private SVMs.
• The algorithm demonstrates minimal error in optimized objective values (less than two).
• DPWSS exhibits higher execution efficiency than standard SVMs, evidenced by fewer iterations.

Conclusions:

• DPWSS is the first algorithm to apply differential privacy to SVM working set selection.
• The proposed method effectively balances data privacy and classification accuracy.
• DPWSS offers a practical and efficient solution for privacy-preserving SVM training.