Anti-money laundering (AML) is garnering increasing attention from governments and media, both domestically and globally. While several cases have been uncovered, the number still does not correspond to the full risk posed by this type of crime in Vietnam, highlighting the necessity for the development of more effective anti-money laundering (AML) systems. Traditional rule-based methods face challenges in terms of adaptability and accuracy when flagging suspicious transactions. Graph Neural Networks (GNNs) have emerged as a highly regarded model due to their ability to learn from graph-structured data, such as society networks, and financial transaction networks. This study experimented with the application of rewiring algorithms such as Batch Ollivier Ricci Flow (BORF) and First-order Spectral Rewiring (FoSR) to address over-smoothing and overs-quashing problems, enhancing the accuracy of GNNs for multi-layer cases. To the best of my knowledge, this is the first study to evaluate BORF on a financial dataset. The results demonstrated that Graph-SAGE, utilizing the Elliptic imbalanced dataset combining both global and local features and employing FoSR optimization, achieved the highest accuracy in most experiments, with F1 score and Recall score outperforming statistic machine learning models. These findings indicate the strong potential for GNNs to be implemented in AML systems.