Related Experiment Videos
Explainable Federated Learning Model for Cross-Institutional Digital Financial Data Collaboration and Risk Control
1Business School, Nantong Institute of Technology; cao15851272136@163.com.
Journal of Visualized Experiments : Jove
|March 23, 2026
Summary
FedRisk enhances financial risk management by enabling secure, explainable federated learning across institutions. This framework improves fraud detection and loan default prediction while ensuring data privacy and regulatory compliance.
Area of Science:
- Financial Technology
- Artificial Intelligence
- Data Privacy
Background:
- Financial data fragmentation and privacy regulations impede collaborative risk management.
- Current federated learning (FL) methods struggle with Non-IID data, interpretability, and compliance, limiting adoption.
- Suboptimal fraud detection (<60% accuracy) and high SME loan default rates (>25%) highlight existing system limitations.
Purpose of the Study:
- To introduce FedRisk, an explainable federated learning framework designed for cross-institutional financial risk management.
- To address the limitations of existing FL solutions in terms of data heterogeneity, interpretability, and privacy compliance.
- To enable secure, transparent, and regulation-compliant risk prediction without sharing raw data.
Main Methods:
- Implemented a cross-institutional data mesh with ontological semantic harmonization for enhanced feature utilization (>85%).
- Developed a hybrid AI model combining interpretable GBDT with monotonic constraints and a global attention-based Transformer.
- Integrated a tripartite privacy mechanism: adaptive differential privacy (ε=5), additive secret sharing, and fairness-aware regularization.
- Incorporated communication-efficient optimizations reducing bandwidth (8.5 MB/round) and training time (85 min).
Main Results:
- FedRisk achieved an AUC-ROC of 0.912 on over 5 million records, closely matching centralized GBDT performance (1.8% lower).
- Reduced performance loss in Non-IID scenarios by 7.6% compared to FedAvg and outperformed FL baselines by 7.3% in AUC-ROC.
- Limited group prediction bias (<0.05) and RAROC loss (<2.1%), while resisting membership inference attacks (AUC=0.58).
- Reduced cross-institutional data transmission by 99%.
Conclusions:
- FedRisk offers a scalable and compliant solution for cross-institutional financial risk prediction.
- The framework significantly improves accuracy in fraud detection and loan default prediction.
- FedRisk enables effective collaborative risk management while preserving data privacy and regulatory adherence.