Privacy-Preserving Cost-Sensitive Learning.

Area of Science:

• Machine Learning
• Data Privacy
• Algorithm Design

Background:

• Increasing use of sensitive personal data necessitates privacy-preserving methods.
• Existing cost-sensitive learning lacks optimal schemes under differential privacy constraints.

Purpose of the Study:

• To develop a unified framework for cost-sensitive learning methods incorporating privacy guarantees.
• To propose novel privacy-preserving algorithms for cost-sensitive classification.

Main Methods:

• Developed a unified framework using regularized empirical risk minimization with weight constants and functions.
• Proposed two privacy-preserving algorithms: output perturbation and objective perturbation.
• Analytically derived privacy-preserving extensions for logistic regression and support vector machines.

Main Results:

• Experimental results demonstrate effective reduction in misclassification cost while meeting privacy requirements.
• Theoretical analysis reveals an interaction between weight functions, privacy levels, and classifier performance.
• The proposed framework successfully integrates cost-sensitive learning with differential privacy.

Conclusions:

• The developed framework offers an optimal approach for cost-sensitive learning under differential privacy.
• The choice of weight functions significantly impacts classifier performance and privacy levels.
• This work provides a foundation for privacy-preserving machine learning in sensitive data applications.