Privacy-Aware Financial Risk Control: A Federated Learning Approach with Differential Privacy Optimization

Abstract

This study proposes a financial risk control and privacy protection method based on federated learning (FL) to address the challenges of traditional centralized risk control models in data silos, privacy leakage risks, and cross-institutional collaboration. Financial risk control relies on a large amount of user transaction, credit score, and market behavior data, but due to privacy regulations (such as GDPR, CCPA), it is difficult for financial institutions to directly share data, resulting in limited generalization of risk control models. Federated learning enables multiple financial institutions to collaboratively optimize risk control models without data leaving the local area through distributed training, effectively protecting data privacy. This study constructs different FL architectures, including horizontal FL, vertical FL, and federated transfer learning, and analyzes their impact on risk assessment models. In addition, we introduce a differential privacy (DP) mechanism to evaluate its impact on model performance (AUC, Precision) while protecting user data. The experiment is verified based on the FICO Credit Score Dataset. The results show that FL performs better than traditional centralized learning methods in risk control tasks, and that appropriately adjusting the DP level can strike a balance between privacy protection and model performance. This study provides a secure and efficient data collaboration solution for financial risk control and lays the foundation for the development of future financial privacy computing technology.