FRAUD-X: An explainable AI-based deep learning model for real-time cloud-driven financial fraud and risk analysis

Abstract

FRAUD-X is an explainable, a cloud-native explainable deep learning framework designed to improve real-time financial fraud detection while ensuring interpretability and regulatory compliance. The primary objective is to develop a scalable Temporal Convolutional Network (TCN)-based model integrated with Explainable Artificial Intelligence (XAI) techniques to reduce false positives and enhance transparency in fraud prediction. The model is evaluated using the Kaggle Financial Fraud Detection dataset containing 284,807 transactions, of which 492 are fraudulent (0.172%), addressing severe class imbalance through SMOTE and weighted loss optimization. The proposed framework combines causal and dilated convolutions for sequential pattern learning with SHAP, Attention Maps, and Saliency Maps for local and global interpretability. Experimental results demonstrate superior performance compared to CNN, LSTM, Transformer, and SVM models, achieving 98.2% accuracy, 94.7% precision, 91.1% recall, and an F1-score of 92.9%, while significantly reducing false positives. The findings suggest that integrating explainable deep learning with cloud-based deployment enhances institutional trust, supports regulatory auditing, and improves operational fraud risk management. The originality of this research lies in the unified integration of TCN-based sequential modelling, multi-layer XAI interpretability, and real-time cloud deployment within a single end-to-end fraud detection architecture. Combined with end-to-end visualization and interpretability features, FRAUD-X empowers fraud analysts to detect threats and understand the reasoning behind alerts. The study concludes with proposals for enhancing explainability and outlines future research directions.