Financial fraud poses a significant and evolving threat in the landscape of electronic transactions. This study presents the design, implementation, and evaluation of an explainable AI (XAI) system for credit card fraud detection. The system integrates a deep neural network (DNN) autoencoder for fraud detection, SHAP (SHapley Additive exPlanations) for model interpretation. Results demonstrate the effectiveness of the DNN in detecting fraudulent transactions and the ability of SHAP to provide insights into influential features. Notably, user feedback revealed a strong preference for LLM-generated narrative explanations over traditional SHAP visualizations. This research contributes to the field of XAI in finance by showcasing a practical approach to enhancing the interpretability and user-friendliness of fraud detection systems, emphasizing the importance of accessible explanations in building user trust, and understanding in AI driven financial decision-making.
Keywords: SHAP (SHapley Additive exPlanations), Financial fraud, Large Language Model (LLM), Fraud Detection System