Data Financial Loan Risk Analysis and Evaluation Based on Federated Learning

ABSTRACT With the rapid economic growth and rising personal consumption demands in our country, bank loans have become a vital means for individuals and enterprises to meet funding needs. This study proposes a bank lending risk assessment system based on federated learning, aiming to reduce financial risks while safeguarding data privacy. The system adopts a horizontal federated learning framework that integrates the FedAug and XGBoost algorithms for performance optimization. Blockchain technology is incorporated to build a decentralized enterprise credit evaluation model, enabling collaborative training via smart contracts. To further strengthen privacy protection, differential privacy techniques are embedded into a PATE model within the horizontal federated learning framework, ensuring the security of both model parameters and local data. Experimental results show that, after optimizing computational resources and convergence speed, the system achieves an average AUC of 95.5%, marking an improvement of about 10% over traditional methods. Comparative analysis also reveals that the Fed‐XGB algorithm outperforms other models in convergence and stability, validating the system's effectiveness.