Privacy Preserving Risk Modeling Across Financial Institutions via Federated Learning with Adaptive Optimization

Abstract

This study presents a privacy-aware risk control method based on federated learning, named FedRisk, which aims to mitigate the long-standing conflict between data isolation and information sharing among financial institutions. By integrating the FedAvg algorithm with differential privacy, the method allows banking, e-commerce, and insurance entities to update model parameters jointly without exposing raw user data. To address distributional discrepancies caused by non-independent and identically distributed (non-IID) data, a dynamic weighting scheme is applied. The approach is validated using real-world data from 820,000 users, covering contract performance, repayment behavior, and credit defaults. Compared with a conventional centralized XGBoost model, FedRisk shows a moderate drop in AUC from 0.874 to 0.861 (approximately 1.5%) but effectively safeguards user privacy. In out-of-bag (OOB) testing, the F1-score improves by 3.7%, suggesting better adaptability to unseen data. Overall, FedRisk provides a practical balance between model performance and privacy preservation in financial risk detection across institutions.