Scholarlink

Federated learning (FL) supports distributed model training among various financial institutions while maintaining sensitive client information intact. As opposed to centralized methods, FL supports credit risk prediction without the exposure of raw data, hence being a fit for privacy-critical finance use cases. FL is still open to adversarial attacks like label-flipping, data poisoning, and Byzantine attacks. This paper introduces a safe and explainable FL framework for credit risk evaluation based on differential privacy, strong aggregation, and a copula-based reputation mechanism to assess node trustworthiness. Explainability is guaranteed using SHAP-based feature attribution for transparent decision-making. Experimental analyses on the Lending Club and German Credit datasets show that the framework supports high prediction accuracy (91.3% accuracy, 0.94 AUC), 28% communication overhead reduction, and model integrity under non-IID and adversarial attacks. These results confirm the framework’s real-world suitability for secure, explainable, and privacy-friendly credit assessment in decentralized finance ecosystems.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Secure and Explainable Federated Learning for Credit Risk Prediction in Decentralized Financial Systems

  • 章节

  • 发表日期：

    2026年

查看全文
  • S. Kiruthika,
  • Naveenbalaji Gowthaman,
  • S. Sathya Bama

摘要

Federated learning (FL) supports distributed model training among various financial institutions while maintaining sensitive client information intact. As opposed to centralized methods, FL supports credit risk prediction without the exposure of raw data, hence being a fit for privacy-critical finance use cases. FL is still open to adversarial attacks like label-flipping, data poisoning, and Byzantine attacks. This paper introduces a safe and explainable FL framework for credit risk evaluation based on differential privacy, strong aggregation, and a copula-based reputation mechanism to assess node trustworthiness. Explainability is guaranteed using SHAP-based feature attribution for transparent decision-making. Experimental analyses on the Lending Club and German Credit datasets show that the framework supports high prediction accuracy (91.3% accuracy, 0.94 AUC), 28% communication overhead reduction, and model integrity under non-IID and adversarial attacks. These results confirm the framework’s real-world suitability for secure, explainable, and privacy-friendly credit assessment in decentralized finance ecosystems.