A Study on the Design of a Cross-Financial Institution Risk Modelling System Based on Federated Learning

The deepening application of artificial intelligence technologies within the financial sector has drawn significant attention to research on cross-institutional intelligent risk control systems. However, traditional centralised architectures struggle to address data privacy protection and inter-institutional collaboration challenges. This study aims to design a cross-financial institution risk modelling system based on federated learning, enabling collaborative modelling without data leaving its domain. The research employs a distributed machine learning framework, combining differential privacy and homomorphic encryption technologies to safeguard data security, while optimising the model training process through an enhanced federated aggregation algorithm. Experimental validation on 380 million data samples from ten financial institutions demonstrates that the system achieves an overall AUC of 0.926, representing a 4.7 percentage point improvement over traditional methods. It supports 2,000 concurrent user accesses with system availability reaching 99.99%. The research indicates that federated learning-based intelligent risk control systems can effectively balance data security and model performance, offering a novel technical pathway for financial technology innovation.