Federated Machine Learning for Privacy-Preserving Credit Risk Assessment Across Distributed Financial Institutions

Abstract

Credit risk models trained at individual financial institutions can suffer from limited observations, heterogeneous borrower populations, and restrictions on sharing sensitive financial records. This paper proposes a cross-silo federated machine learning architecture for probability-of-default estimation across distributed lenders. The design combines locally standardized borrower features, secure aggregation, differential privacy accounting, institution-aware calibration, and fairness-aware evaluation. We formulate a weighted federated objective and a reproducible experimental protocol covering temporal validation, non-identically distributed clients, membership-inference exposure, predictive utility, calibration, and operational cost. Because no institution-level data or executed experiment was supplied, all numerical tables and charts in the results section are explicitly synthetic scenario illustrations, not measured findings. The framework provides an auditable blueprint for empirical implementation while identifying the tradeoffs among discrimination, privacy, calibration, fairness, and communication efficiency.