Hybrid Framework of Graph Neural Networks, Federated Learning and Explainable Artificial Intelligence for Real-Time Fraud Detection and Anti-Money Laundering

Abstract

The increased sophistication of fraudulent schemes and money laundering techniques is placing significant pressure on conventional fraud detection systems and models. This study presents a hybrid framework that integrates Graph Neural Networks, Federated Learning and Explainable Artificial Intelligence for real-time fraud detection and anti-money laundering in digital financial systems. The framework uses graph neural networks to learn heterogeneous financial transaction graphs which helps to uncover organized fraud committed by fraudsters. Federated learning was leveraged in this study to collaboratively train global graph neural network fraud detection model while protecting the privacy of the contributing banks using differential privacy. In addition, the framework was supplemented with explainable artificial intelligence, namely SHAP and GNNExplainer to enhance the model’s interpretability. The proposed framework was benchmarked against conventional machine learning algorithms, deep learning and graph neural networks on the SynthAML dataset which contains more than 16 million transactions in a large-scale graph. The evaluation results revealed that the framework achieved an accuracy of 97.6%, precision of 95.7%, recall of 92.6%, F1-score of 94.1% and ROC-AUC of 0.984. The framework had better performance compared to conventional fraud detection models, deep learning and graph neural networks. In federated learning experiments, the proposed graph neural networks converged within fifty communication rounds while achieving reasonable performance with minimal communication overhead. Specifically, the performance decreased by only 0.8% when applying differential privacy for the model. Additionally, the explainability component of the framework achieved a faithfulness score of 0.913. In conclusion, the hybrid framework outperformed conventional fraud detection models, deep learning and graph neural networks. The framework can be leveraged for fraud detection and anti-money laundering detection in digital financial systems.