Explainable artificial intelligence models for detecting suspicious bank transactions

Detecting financial crime is a complex challenge due to evolving criminal strategies and fragmented detection systems, particularly in the areas of money laundering and fraud. While it is easy to implement, traditional rule-based approaches lack adaptability to new threats, and machine learning models, though more effective, often function as opaque "black boxes," limiting their practical use in regulated domains like banking, where interpretability and accountability are essential. This research presents a novel framework that combines intrinsic and post-hoc XAI techniques to detect suspicious bank transactions. Intrinsic methods provide model-inherent transparency, while post-hoc methods offer behavior-level explanations, enabling robust cross-verification of the model’s decision logic. This dual-layered explainability supports both global understanding of the model and local interpretability for individual predictions. Experimental evaluation on synthetic datasets shows that the proposed framework achieves promising results, with F1 scores ranging from 0.54 to 0.63 for anti-money laundering (AML) detection and 0.66 to 0.78 for fraud detection, depending on the underlying model architecture. The use of explainability tools also enables the identification of key features shared across fraud and AML cases, revealing structural similarities. These insights demonstrate that a unified model can effectively capture both crime types, offering practical benefits for integrated financial crime risk management. By addressing the information silos caused by different regulatory frameworks and departmental responsibilities, this work supports the development of joint investigative systems and facilitates better communication between compliance teams. In addition, the explainable nature of the models helps satisfy regulatory scrutiny while improving investigator trust and decision support. The proposed approach contributes to the growing field of interpretable machine learning by providing a scalable, auditable, and regulator-friendly solution for detecting complex financial crimes.