Graph neural network-based anomaly detection in financial transactions associated with money laundering

The money laundering is a global threat, compromising the integrity of the financial system and risking the stability of the global economy. This work proposes the use of complex network techniques and presents a methodology to detect anomalies in financial transactions of individuals under investigation for suspected money laundering. The methodology involves creating various financial indicators, such as the average, sum of transaction values, and the number of transactions sent and received by each bank account. In this context, the account number represents the node in the directed graph. The Unifying Local Outlier Detection Methods via Graph Neural Networks (LUNAR) algorithm was used to recognize patterns in financial transactions and identify anomalies. The results highlight the model's effectiveness, with Silhouette score and Davies-Bouldin Index metrics of 1.59 and 0.83 achieved on the test set, respectively. This indicates that groups of anomalous and normal accounts are well represented in terms of similarity and dissimilarity. The results are promising and may assist in investigations by helping to identify potential groups of individuals involved in illicit activities, such as drug and arms trafficking, fraud, and scams.