Fairness and privacy-preserving credit evaluation using machine learning techniques

Brian Cu

Credit evaluation is vital in the banking sector globally, including Vietnam’s banking sector, to shape lending decisions and financial stability. However, it faces significant challenges regarding fairness and privacy. Biased decisions based on sensitive attributes, such as gender or ethnicity, can result in discrimination. For instance, studies by the State Bank of Vietnam have highlighted instances of gender disparities, where male applicants often receive preferential treatment. Additionally, privacy concerns arise due to the sensitive nature of financial data. Our research addresses these challenges by proposing a novel approach to enhance transparency, fairness, and privacy. Leveraging Support Vector Machines with a dataset called Default, which is renowned for their realistic representation of credit assessment scenarios, including various demographic and financial attributes of borrowers, we aim to build a fairness-aware and privacy-preserving credit evaluation system. This solution integrates messaging techniques for fairness enhancement and applies Differential Privacy techniques for privacy protection, chosen for their simplicity and effectiveness in promoting fairness and privacy. Our study reveals a significant reduction in the True Positive Rate (TPR) difference between male and female applicants, from 13.5% to 4.7%, with fairness processing technique. After applying Differential Privacy techniques, our model maintains this improvement, achieving a TPR difference of approximately 4.7% at the moderate level of privacy (denoted by ϵ) withϵ = 10. This illustrates that our model can uphold satisfactory fairness and privacy of users while maintaining satisfactory performance. By promoting fairness and privacy in credit assessments, our work contributes to creating a more inclusive and equitable financial system, fostering trust among borrowers and lenders. The impact of our research can extend to areas where underprivileged people need to be protected in various circumstances, such as recruitment, legal judgment etc. Moving forward, we plan to apply fairness and privacy techniques beyond margin classifiers to deep learning in our future research.

Fairness and privacy-preserving credit evaluation using machine learning techniques

Files and links (1)

Abstract

Details

Metrics