FedRisk A Federated Learning Framework for Multi-institutional Financial Risk Assessment on Cloud Platforms
Xiaoxiao Jiang
Abstract
This paper introduces FedRisk, a novel federated learning framework designed for multi-institutional financial risk assessment on cloud platforms. Traditional financial risk management systems face significant challenges in cross-institutional contexts, including data silos, privacy concerns, and computational inefficiencies. FedRisk addresses these challenges by enabling collaborative model building while preserving data privacy and security. The framework implements a distributed approach where institutions train models locally using proprietary data, sharing only model parameters rather than raw data. We integrate knowledge graph technology with a specialized parameter aggregation strategy that accounts for data heterogeneity across participating institutions. Experimental results using financial data from 70 companies demonstrate that FedRisk significantly outperforms both centralized approaches and existing federated learning solutions, achieving 93.7% accuracy and 88.3% recall in financial crisis prediction. Under severe data heterogeneity conditions, FedRisk exhibits minimal performance degradation (12.3%) compared to traditional federated averaging (26.8%). Additionally, the framework demonstrates superior communication efficiency, requiring only 0.16-0.18 GB of total data transfer, a 6-7× improvement over baseline methods. FedRisk provides a comprehensive solution for privacy-preserving, efficient, and accurate financial risk assessment across institutional boundaries.