Real-Time Cross-Border Anti-Money Laundering (AML) Detection Using Graph Neural Networks and Machine Learning

Abstract

Cross-border money laundering involves complex, rapidly changing transaction relationships that are difficult to identify using conventional rule-based and isolated transaction-level machine learning systems. This paper proposes a real-time Anti-Money Laundering (AML) detection framework that integrates Graph Neural Networks (GNNs) with machine learning to model financial transactions as dynamic, heterogeneous graphs. Customers, accounts, financial institutions, merchants, devices, jurisdictions, and transactions are represented as interconnected nodes and edges, enabling the model to capture transactional dependencies, circular fund movements, layering patterns, intermediary accounts, and cross-border risk propagation. A temporal graph-learning mechanism is incorporated to process streaming transactions and identify emerging suspicious structures with low detection latency. Supervised machine learning is combined with graph-based embeddings to improve classification under severe class imbalance, while anomaly detection supports the identification of previously unseen laundering behaviours. The proposed framework further incorporates risk scoring, explainable graph patterns, and privacy-aware cross-institutional information exchange. The resulting architecture provides an integrated foundation for scalable, adaptive, and real-time AML surveillance across international financial networks.