Privacy-Preserving Techniques in Credit Risk Assessment: A Comparative Analysis of Differential Privacy, Federated Learning, and Homomorphic Encryption

The proliferation of machine learning in financial services has intensified concerns regarding consumer data privacy during credit risk assessment. This paper presents a comparative study of three privacy-preserving paradigms relevant to credit risk assessment: differential privacy, federated learning, and homomorphic encryption. Through empirical evaluation on a large-scale, representative credit dataset across multiple privacy-preserving configurations, we examine the privacy-utility trade-offs inherent in each approach. Our experimental framework assesses prediction accuracy, computational overhead, and privacy guarantees across multiple configurations. Results demonstrate that differential privacy with epsilon=8.65 achieves a balanced privacy-utility tradeoff with competitive discriminative performance, while homomorphic encryption preserves near-baseline predictive performance during encrypted inference. Federated learning enables collaborative model training across institutions without sharing raw data. The analysis reveals that privacy parameter selection critically impacts model utility, with differential privacy offering quantifiable privacy budgets and homomorphic encryption providing cryptographic security guarantees. These findings provide actionable guidance for financial institutions navigating regulatory compliance requirements while maintaining effective risk management capabilities.