Detecting Financial Fraud in Real-Time Transactions Using Graph Neural Networks and Anomaly Detection

Abstract

The exponential growth of digital financial services has amplified the risk and complexity of fraud in real-time transactional systems. Traditional rule-based or statistical approaches are often inadequate for detecting evolving and covert fraudulent behaviors embedded within large-scale financial networks. This paper proposes a novel, data-driven framework that leverages Graph Neural Networks (GNNs) combined with unsupervised anomaly detection to identify fraudulent activity in real-time transaction streams. By modeling financial transactions as a dynamic graph, where nodes represent users/accounts and edges represent transactions, the system captures the intricate relational patterns and dependencies among entities. A GNN is then trained to learn latent representations of nodes and edges, which are subsequently analyzed using density-based anomaly scoring techniques such as Isolation Forest and Local Outlier Factor (LOF). Our experimental results, conducted on publicly available and simulated financial datasets, demonstrate that the proposed hybrid model significantly outperforms baseline methods in terms of detection accuracy, precision, and false positive rates. Furthermore, the system offers real-time inference capabilities, making it highly applicable for deployment in fraud monitoring engines of banks, fintech platforms, and payment gateways. This study establishes the effectiveness of graph-based deep learning and unsupervised anomaly detection as a unified solution for modern financial fraud prevention.