Scholarlink

<p>Money laundering poses a major threat to global financial systems, yet efforts to detect illicit transactions are hindered by strict data privacy regulations and institutional silos. This study introduces a privacy-preserving federated learning (FL) framework for collaborative money laundering detection across financial institutions, without sharing sensitive data. The proposed system integrates Random Forest classifiers with differential privacy (DP) and secure aggregation to ensure data confidentiality and regulatory compliance. We address class imbalance using random undersampling at the client level and demonstrate that our model achieves up to 85.71% accuracy on the IBM Anti-Money Laundering dataset. Experimental results highlight the trade-offs between privacy budgets and model utility, with secure model sharing enabled through homomorphic encryption. The framework is communication-efficient, suitable for deployment on resource-constrained clients, and robust to varying privacy configurations. Our results show that privacy-preserving FL can enable effective, collaborative anti-money laundering solutions in real-world financial environments.</p>

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

A Privacy-Preserving Federated Learning Model for Money Laundering Detection

  • 期刊论文

  • 发表日期：

    2025年6月25日

查看全文
  • Huu Huong Xuan Nguyen,
  • Tran Khanh Dang

摘要

Money laundering poses a major threat to global financial systems, yet efforts to detect illicit transactions are hindered by strict data privacy regulations and institutional silos. This study introduces a privacy-preserving federated learning (FL) framework for collaborative money laundering detection across financial institutions, without sharing sensitive data. The proposed system integrates Random Forest classifiers with differential privacy (DP) and secure aggregation to ensure data confidentiality and regulatory compliance. We address class imbalance using random undersampling at the client level and demonstrate that our model achieves up to 85.71% accuracy on the IBM Anti-Money Laundering dataset. Experimental results highlight the trade-offs between privacy budgets and model utility, with secure model sharing enabled through homomorphic encryption. The framework is communication-efficient, suitable for deployment on resource-constrained clients, and robust to varying privacy configurations. Our results show that privacy-preserving FL can enable effective, collaborative anti-money laundering solutions in real-world financial environments.