Federated Learning for Credit Risk Assessment in Distributed Financial Systems using BayesShield with Homomorphic Encryption

Accurate credit risk assessment is essential for maintaining the stability and dependability of lending systems in the rapidly changing world of data-driven finance. However, conventional machine learning methods rely heavily on centralized data collection, which introduces serious concerns related to data privacy, security breaches, and regulatory non-compliance. These challenges are particularly severe in the financial domain, where institutions are reluctant to share sensitive customer data. To overcome these limitations, this paper presents BayesShield, a secure and scalable federated learning framework specifically designed for credit risk assessment. BayesShield combines the interpretability and speed of Naïve Bayes with the high accuracy and scalability of LightGBM, offering a hybrid ensemble classifier suited for heterogeneous, non-IID financial data. To maintain privacy while updating models, Homomorphic Encryption (HE) is used with the TenSEAL library, such that secure aggregation of parameters without raw data exposure is possible. Experiments conducted on credit risk dataset demonstrate that the proposed method achieves better accuracy (93%) and F1-score when compared to baseline classifiers like Naïve Bayes, XGBoost, LightGBM, and Random Forest. The findings confirm BayesShield's efficacy in maintaining data privacy while providing strong, interpretable, and high-performing credit risk predictions in non-IID federated settings.