Exploration of Federated Multi-Task Learning Models for Secure Cross-Institutional Credit Risk Assessment under Privacy Constraints

As financial institutions increasingly collaborate for more robust credit risk assessment, privacy-preserving machine learning techniques have become essential. This paper investigates Federated Multi-Task Learning (FMTL) as a scalable, privacy-preserving framework for credit risk modeling across heterogeneous institutions. Our proposed approach enables multiple financial institutions to jointly train models without sharing raw data, leveraging the task-specific nuances of each institution’s portfolio. We introduce novel optimization strategies that incorporate differential privacy and secure aggregation protocols. Extensive experiments on synthetic and real-world financial datasets demonstrate improved prediction accuracy and fairness compared to traditional federated and centralized learning baselines. The findings support the viability of FMTL as a regulatory-compliant, data-secure solution for inter-institutional credit modeling