Hybrid Deep Learning and Synthetic Transaction Simulation Framework for RealTime Financial Fraud Detection using VAE and BiLSTM

Abstract

Financial fraud has become increasingly sophisticated with the rapid growth of digital transactions, posing serious threats to banking systems and e-commerce platforms. Traditional rule-based systems are no longer sufficient to detect complex fraud patterns due to their inability to adapt to evolving attack strategies. This paper proposes a hybrid deep learning-based fraud detection framework that integrates Variation Auto encoder (VAE) and Bidirectional Long Short-Term Memory (BiLSTM) networks with a real-time web-based monitoring system built using Django. The proposed system focuses on detecting fraudulent financial transactions by combining anomaly detection and sequential behavior analysis. The VAE model is employed to learn the normal transaction distribution and identify anomalies based on reconstruction error. Transactions that significantly deviate from learned patterns are assigned higher anomaly scores. Simultaneously, the BiLSTM model captures temporal dependencies in transaction sequences, enabling the system to identify suspicious behavioral patterns over time. To address the scarcity of labeled fraud datasets, the system incorporates a synthetic transaction generation engine. This module simulates realistic transaction scenarios by introducing controlled fraud patterns based on transaction amount, category, velocity, and time intervals. The generated data is then evaluated using the trained models to produce risk scores, enhancing the robustness and scalability of the system. A weighted ensemble approach is used to combine outputs from both models, producing a unified risk score that determines whether a transaction should be flagged as potentially fraudulent. The system dynamically updates performance metrics such as accuracy, precision, recall, and model losses, providing real-time insights into model performance through an interactive dashboard. The Django-based web interface offers multiple modules including transaction streaming, threat hunting, system health monitoring, and configuration management. It enables users to visualize transaction data, monitor suspicious activities, and evaluate system performance in real time. Experimental results demonstrate that the hybrid model significantly improves detection accuracy while reducing false positives compared to standalone models. The integration of synthetic data generation further enhances the model’s ability to generalize to unseen fraud patterns. Overall, the proposed framework provides a scalable, adaptive, and intelligent solution for real-time financial fraud detection, making it highly suitable for modern digital transaction ecosystems.