Explainable artificial intelligence models for detecting suspicious banking transactions

Kumar K Narasapuram

Financial crimes, particularly money laundering and fraud, represent significant global threats, with evolving techniques and increasing sophistication from criminals. Financial institutions tasked with identifying and mitigating these crimes have relied on rule-based systems to detect suspicious transactions. However, these static approaches have become increasingly ineffective as they fail to keep pace with the rapid advancements in criminal methods. In addressing this challenge, this thesis proposes the integration of Explainable Artificial Intelligence (XAI) into machine learning models to enhance the detection of suspicious banking transactions, specifically those that may involve both money laundering and fraud. Traditional methods for detecting financial crimes have largely focused on either fraud or money laundering in isolation, leading to fragmented approaches that overlook the overlap between these activities. Additionally, while machine learning techniques have been applied to improve detection accuracy, most models function as "black boxes," making their decision-making processes difficult to interpret and trust. This lack of transparency is particularly problematic in high-stakes sectors like banking, where regulatory bodies and stakeholders demand clear explanations for automated decisions. The primary contribution of this research lies in bridging these gaps by developing a framework that simultaneously detects both fraud and money laundering while also providing interpretable insights into the machine learning models used. The study employs a combination of post-hoc and intrinsic XAI methods to explain the decision-making process of the models, identifying the top contributing features and calculating correlation coefficients that highlight the relationships between key variables. These insights not only improve model transparency but also allow investigators to better understand how certain transactions are flagged as suspicious. In the first phase of the research, traditional machine learning models, including Random Forest (RF), Support Vector Machine (SVM), and Gradient Boosting (GB), are applied to identify suspicious transactions. These models are tested separately for fraud detection, anti-money laundering (AML) detection, and the combined detection of both. The second phase focuses on applying XAI techniques to the results, allowing for a detailed interpretation of the model’s behaviour and the factors influencing its decisions. By identifying common features across fraud and AML transactions, the study demonstrates that a unified model can effectively capture both types of financial crimes, addressing a significant gap in the current literature. The findings of this thesis have important implications for the financial industry. They suggest that financial institutions should consider integrating fraud and AML detection systems, rather than treating them as separate processes, to improve the overall effectiveness of their risk management strategies. Furthermore, the use of XAI enhances the interpretability and trustworthiness of AI-driven systems, providing clear and understandable justifications for automated decisions. This transparency is crucial for regulatory compliance and for fostering trust among stakeholders, including auditors and financial crime investigators. Overall, this thesis not only advances state-of-the-art financial crime detection by combining fraud and money laundering into a single, explainable framework but also sets the stage for future research in the application of XAI to other domains requiring both high accuracy and interpretability in machine learning models. The research underscores the necessity of explainability in AI systems and advocates for a shift towards more transparent and unified approaches to financial crime detection.

Explainable artificial intelligence models for detecting suspicious banking transactions

Files and links (1)

Abstract

Details

Metrics