Differentially Private Singular Value Decomposition for Training Support Vector Machines.

Area of Science:

• Machine Learning
• Data Privacy
• Dimensionality Reduction

Background:

• Support Vector Machines (SVMs) are effective for classification but risk privacy when using sensitive training data.
• Principal Component Analysis (PCA) reduces dimensionality by projecting data into a lower-dimensional subspace.
• Singular Value Decomposition (SVD) is a method for PCA that avoids covariance matrix computation, unlike Eigenvalue Decomposition (EVD).

Purpose of the Study:

• To develop a novel differentially private Singular Value Decomposition (DPSVD) algorithm.
• To safeguard sensitive information within training datasets for SVM classifiers.
• To enable training of SVM models using projected data without revealing original instance privacy.

Main Methods:

• The study proposes a new differentially private Singular Value Decomposition (DPSVD) algorithm.
• DPSVD generates private singular vectors for data projection.
• The theoretical differential privacy of DPSVD was formally proven.

Main Results:

• The DPSVD algorithm successfully generates private singular vectors.
• Projected instances in the singular subspace are suitable for training SVM classifiers.
• Experimental results demonstrate higher accuracy and stability compared to existing private PCA methods for SVM training.

Conclusions:

• The proposed DPSVD algorithm effectively protects data privacy in SVM classification.
• DPSVD offers a robust solution for privacy-preserving machine learning applications.
• The method achieves superior performance in terms of accuracy and stability on real-world datasets.