Adaptive Explainable Multi-Agent Fraud Detection Framework Using Large Language Models and Hybrid Machine Learning for Mobile Money Payment Systems

The rapid adoption of Mobile Money Payment Systems (MMPS) has increased financial inclusion, particularly in developing countries such as India, China, Brazil and several African countries. But the increasing use of digital payment platforms has also resulted in the increasing of sophisticated fraud attacks such as SMS phishing, account impersonation, transaction manipulation, and social engineering. The existing fraud detection techniques are largely based on static machine learning and deep learning models that are not able to adapt to changing fraud patterns, have limited interpretability and do not utilize unstructured text. In this paper, we have developed an Adaptive Explainable Multi-Agent Fraud Detection Framework (AEMAF) which combines LLMs, hybrid machine learning and Explainable Artificial Intelligence (XAI) for the fraud detection of MMPS. The framework uses smart agents in SMS semantic analysis, transaction risk assessment, behavioral anomaly detection, device profiling, decision fusion and explanation generation to identify and assess fraud risk and their corresponding behaviors. A multimodal feature fusion module combines these heterogeneous data sources to improve detection accuracy. A feedback system based on an adaptive algorithm continuously updates the detection model on verified fraud cases to protect from concept drift. SHapley Additive exPlanations (SHAP) based on LLM natural language explanations allows the fraud detection system to be transparent and interpretable. The proposed framework will improve fraud detection accuracy, reduce false positives, improve model adaptability and be more secure and transparent for mobile financial services.