Enhancing Anti-Money Laundering Detection with Self-Attention Graph Neural Networks

Money laundering remains a significant global issue, undermining financial stability and security. This study introduces a Self-Attention-GNN Model enhanced with a self-attention mechanism to improve the detection of money laundering activities in a large, imbalanced dataset of financial transactions. The dataset, covering 97 days and including approximately 180 million transactions, contains 223,000 labeled laundering cases. By representing financial transactions as a graph—where entities such as accounts and banks are nodes, and transactions are edges—the model captures intricate relational and structural dependencies within the transaction network. The addition of the self-attention mechanism enables the model to dynamically adjust feature aggregation, focusing on the most relevant nodes and edges, which significantly improves the model’s ability to identify laundering activities. Despite the challenges posed by class imbalance, the model achieves robust performance in detecting illicit transactions while reducing false positives. The paper also discusses potential strategies for further optimizing precision and recall, such as advanced graph architectures, oversampling methods, and enhanced node embedding techniques. Overall, this research highlights the power of graph-based deep learning approaches for anti-money laundering (AML) applications, demonstrating how structural and relational dependencies within financial networks can be leveraged to enhance detection accuracy.