AI Driven Fraud Detection Models in Financial Networks, Cybercrime, Digital Security

Abstract

Financial fraud detection is a critical challenge in modern banking and fintech ecosystems. This paper proposes a comprehensive deep learning framework employing Deep Neural Networks (DNN), Convolutional Neural Networks (CNN), and Long Short-Term Memory (LSTM) networks for robust fraud detection on the Zenodo financial transaction dataset. Extensive feature engineering introduces five domain-specific features — Risk_Score, Liquidity_Index, Profitability_Index, RiskLiquidity_Ratio, and ProfitRisk_Ratio — expanding the feature space to 27 dimensions. Class imbalance is addressed via SMOTEENN, a hybrid resampling technique combining SMOTE oversampling with Edited Nearest Neighbour noise removal. All models are rigorously evaluated using 5-fold stratified cross-validation. The DNN achieves test accuracy of 95.50% with AUC 97.83%; the CNN achieves 94.59% accuracy with AUC 96.72%; Explainability is provided through SHAP global feature importance and LIME local explanation analyses. Experimental results demonstrate that all three proposed models significantly outperform conventional machine learning baselines, establishing a reliable and interpretable automated fraud detection pipeline.