Financial fraud detection presents a critical challenge: balancing high model accuracy with stringent data privacy regulations such as GDPR and CCPA. Centralized machine learning approaches, which require pooled transaction data, pose significant privacy risks, while institution-specific models suffer from data scarcity. We propose a privacy-preserving framework for fraud detection using Horizontal Federated Learning (HFL). Our study compares three paradigms: (i) global centralized models, (ii) partially isolated models, and (iii) HFL, trained using FedAvg with Flower framework. Experiments on the BAF datasets simulate real-world fraud detection scenarios, evaluating key performance metrics, including ROC-AUC, F1-Score, and time efficiency. The results highlight trade-offs between data privacy, model performance, and generalization ability, demonstrating that Federated Learning is a viable alternative that effectively balances security, efficiency, and predictive performance in financial fraud detection.