Transformer based FinBERT framework with explainable artificial intelligence for money laundering detection

Abstract

The increasing complexity of suspicious financial activities has reduced the effectiveness of conventional rule-based monitoring for modern anti-money laundering (AML) applications. This study proposes FinBERT-AML, a transformer-based framework fine-tuned on transaction feature tokens to learn complex dependencies across heterogeneous financial attributes. Through self-attention-based representation learning, the framework jointly models categorical and continuous transaction characteristics while overcoming limitations associated with conventional machine learning and sequential architectures. Experimental evaluation on the SAML-D dataset demonstrates that the proposed leakage-free configuration, with the target-dependent Laundering_type attribute excluded, achieves an accuracy of 92%, outperforming conventional machine learning, deep learning, and tabular transformer baselines. Information-gain analysis is employed to assess feature relevance, while LIME and SHAP provide complementary local and feature-attribution explanations of individual AML predictions. Statistical significance testing further evaluates the reliability of the observed performance improvements.