Hybrid BiLSTM-Autoencoder Framework with Federated Learning for Intelligent Credit Card Fraud Detection

Abstract

The rapid expansion of digital payment systems has significantly increased the complexity and volume of financial transactions, leading to more sophisticated credit card fraud patterns that are difficult to detect using conventional approaches. This study proposes a hybrid fraud detection framework that integrates Bidirectional Long Short-Term Memory (BiLSTM), Autoencoder, and Federated Learning (FL) to enhance detection performance while preserving data privacy. The BiLSTM component captures temporal dependencies in transaction sequences by analyzing user behavior in both directions, enabling more accurate identification of irregular patterns. The autoencoder module functions as an unsupervised anomaly detector by learning representations of normal transactions and identifying deviations through reconstruction errors. To address data privacy constraints, the proposed model is deployed within a federated learning environment, allowing multiple institutions to collaboratively train a global model without sharing sensitive customer data. Experimental evaluation on benchmark datasets demonstrates that the proposed framework achieves superior performance over traditional machine learning and standalone deep learning models, particularly in precision, recall, and overall classification stability. The model effectively handles class imbalance and detects both known and previously unseen fraud patterns. Furthermore, the integration of federated learning enhances generalization by leveraging distributed data sources while maintaining strict confidentiality. This study contributes a scalable, privacy-preserving, and high-accuracy solution for real-world financial fraud detection, supporting secure collaboration across institutions and aligning with modern regulatory requirements.