GAMLNet: a graph based framework for the detection of money laundering

The accuracy of classification algorithms in detecting fraudulent financial activity is critical in assisting human analysts in the task of preventing financial crime. We consider financial transactions in the form of a directed graph, and propose a Graph Neural Network (GNN) model for the identification of money laundering activity. Our method generates a set of structurally aware and statistically significant features for each graph node, and utilizes them as input to the GNN classifier, that comprises of the combination of the layers of two recently proposed message passing architectures. The effectiveness of our approach is demonstrated in experiments with synthetic data that simulate real-world behaviour, and are infused with seven anomalous money laundering topologies. The accuracy of our method is consistently higher than that of other GNNs and tree-based classification methods over datasets of increasing size and increasing imbalance between the fraudulent and benign classes.