The rapid expansion of digital payment systems has been matched by a rise in fraudulent activities, emphasizing the importance of secure and intelligent detection methods. This study proposes a federated learning-based weighted ensemble framework that improves fraud identification accuracy while maintaining confidentiality of transaction records. Rather than sharing sensitive data, individual clients develop local models using XGBoost, CatBoost, and AdaBoost, transmitting only model parameters for safe aggregation. The optimal ensemble weights are calculated using Non-Negative Least Squares (NNLS), resulting in contributions of XGBoost (0.0516), CatBoost (0.9480), and AdaBoost (0.0004), highlighting CatBoost’s primary influence. Tests on the OpenML Credit Card Fraud Detection dataset (284,807 transactions, 31 features) achieved 99.95% accuracy, 0.9691 ROC-AUC, 0.9180 precision, 0.7568 recall, and 0.8296 F1score, surpassing all individual base models. This methodology provides a viable solution for financial organizations, offering high detection performance alongside privacy protection and scalability in real-world digital payment environments.