Credit card fraud detection relies on machine learning models trained on large and diverse datasets. However, stringent privacy regulations such as GDPR and PSD2 restrict centralized data collection, limiting the applicability of traditional approaches. Federated Learning (FL) provides a decentralized solution that allows institutions to collaboratively train models without exposing their raw data. In this work, we propose a novel FL framework for fraud detection and conduct a comparative evaluation of Logistic Regression (LR), Random Forest (RF), and Extreme Gradient Boosting (XGBoost) under conditions of severe class imbalance. Our experiments reveal that Federated XGBoost consistently delivers state-of-the-art performance, achieving an AUROC of 0.998 and an AUPRC of 0.918, representing a 176-fold improvement over a random classifier. It also achieves the highest F1-score (0.845), striking a balance between precision (91.4%) and recall (78.5%). By contrast, Federated RF demonstrates competitive but slightly weaker results (AUROC: 0.985), while Federated LR proves insufficient for this task (AUPRC: 0.182). Notably, Federated XGBoost approaches the accuracy of centralized benchmarks, demonstrating that FL can preserve near-parity in predictive power without requiring data pooling. However, this superior performance comes at a cost: significantly higher computational complexity and communication overhead. Training XGBoost across distributed clients requires substantially more time and bandwidth than lightweight models such as LR. These findings offer clear implications for financial institutions. While simpler models offer efficiency and scalability, ensemble-based approaches, particularly XGBoost, provide the predictive strength necessary for robust fraud detection. Federated Learning thus emerges as a promising paradigm that balances privacy, efficiency, and accuracy, enabling collaborative intelligence in a privacy-preserving manner.