Accurate and efficient detection of fraudulent financial transactions is essential for ensuring security, trust, and stability in modern digital financial systems. However, this task remains challenging due to highly imbalanced datasets, continuously evolving fraud strategies, heterogeneous data distributions across different financial institutions, and strict privacy constraints in distributed environments. To address these challenges, this research proposes a federated learning-based hybrid deep learning model for fraud detection across distributed rural financial environments. The evaluation was conducted using the Credit Card Fraud Detection Dataset (CCFD), which comprises 284,807 financial transactions with 492 fraudulent transactions. This dataset is severely skewed between valid and fraudulent classes. To mitigate class imbalance, the Synthetic Minority Oversampling Technique (SMOTE) was applied during preprocessing to improve the representation of minority fraud samples and enhance model learning capability. The proposed Pity Beetle Algorithm- driven Federated-tuned Recurrent Neural Networks (PBA-FederatedRNN) model integrates Recurrent Neural Networks (RNNs) for sequential transaction behavior learning and the Pity Beetle Algorithm (PBA) for optimized parameter tuning and faster convergence. Federated learning ensures robust data security and privacy preservation by allowing several financial nodes to cooperatively train a global model without exchanging raw data. The model was implemented using Python with TensorFlow/PyTorch in a distributed simulation environment. According to experimental results, the proposed model outperforms existing centralized and federated approaches with 97.8% recall, 98.0% F1-score, 98.7% accuracy, and 98.2% precision. With all factors considered, the proposed PBA-FederatedRNN model offers a flexible, dependable, and highly efficient fraud detection solution in distributed financial systems that protect privacy.
Keywords: Fraud Detection, distributed systems, federated learning, hybrid model, privacy preservation, rural environments, secure aggregation
