Privacy-Preserving Federated Learning Models for Multi-Bank Credit Risk Assessment

Accurate credit risk assessment is essential for maintaining financial stability, yet collaborative modeling across banks is severely constrained by data privacy regulations and competitive concerns. This paper proposes a Privacy-Preserving Federated Learning (PPFL) framework for multi-bank credit risk assessment, enabling financial institutions to jointly train predictive models without sharing raw customer data. The proposed framework integrates federated averaging, secure aggregation, and differential privacy to ensure confidentiality of sensitive financial information while maintaining high predictive performance. Experiments are conducted on both real-world and benchmark credit datasets, partitioned to simulate a cross-bank non-IID environment. The proposed PPFL model achieves an AUC-ROC of 0.86, which is comparable to the centralized model (0.88) and significantly outperforms standalone local bank models (0.79 on average). With differential privacy enabled at a privacy budget of ε = 1.0, the model experiences only a 2.1% reduction in AUC, demonstrating a favorable trade-off between privacy and utility. Secure aggregation successfully prevents leakage of individual bank updates, while communication overhead increases by less than 18% compared to standard federated learning. The results confirm that privacy-preserving federated learning can deliver robust, regulation-compliant, and high-accuracy credit risk prediction, making it a practical solution for collaborative analytics in multi-bank financial ecosystems.