A Federated Ensemble Learning Framework for Distributed Fraud Detection

With the rapid evolution of digital payment systems and financial services, the number of fraudulent transactions is increasing, and risks are becoming increasingly critical. Although several fraud detection approaches have been proposed, they remain hampered by certain limitations, including confidentiality constraints on cross-institutional data sharing and class imbalance in fraud datasets. To address these challenges, we propose a new hybrid fraud detection framework that integrates federated learning with ensemble learning, enabling collaborative and efficient model training across distributed financial institutions without sharing raw data. The framework leverages heterogeneous machine learning models (XGBoost, CatBoost, and MLP) trained distributedly in a federated architecture, coordinated by a central aggregation server. The three federated models are combined using an ensemble learning method to improve predictive performance and generate more accurate decisions. This solution can help to effectively detect fraud in distributed environments while reducing the need for direct data sharing. Experimental results demonstrate that the proposed federated framework offers competitive performance in terms of recall, F1-score, and AUC-PR similar to, or even superior to, centralized models in certain federated configurations.