HAMLET: A Transformer Based Approach for Money Laundering Detection

Abstract

Money laundering has damaging economic, security, and social consequences, fueling criminal activities like terrorism, human and drug trafficking. Recent technological advancements have increased the complexity of laundering operations, prompting financial institutions to use more advanced Anti Money Laundering (AML) techniques. In particular, machine learning-based transaction monitoring can complement traditional rule-based systems, lowering the number of false positives and the need to manually review fraud alerts. In this paper, we present HAMLET, a scalable Deep Learning model for analyzing financial transactions and detecting money laundering patterns. HAMLET employs a hierarchical transformer enforcing an attention mechanism at the transaction and sequence level. By combining the two different levels, HAMLET can identify complex money laundering operations carried out through subsequent transactions. We experimentally evaluate HAMLET on a synthetic dataset simulating clients trading on international capital markets, showing that HAMLET outperforms state-of-the-art solutions in detecting fraudulent transactions. In the experiments, HAMLET achieves $$99\%$$ precision in a binary classification scenario and up to $$95\%$$ in a multi-class scenario with 5 different money laundering schemes. Lastly, we investigate the proposed model’s interpretability through a proxy method.