Detecting Fraudulent Online Transactions Using Deep Neural Networks with SMOTE Oversampling and Explainable AI Techniques

Online banking fraud has emerged as a critical challenge in the digital financial ecosystem. With millions of transactions processed daily, traditional rule-based detection mechanisms are insufficient to identify evolving fraudulent patterns. This study applies machine learning and deep learning techniques — specifically an Artificial Neural Network (ANN) — to classify transactions from a publicly available Kaggle dataset as fraudulent or legitimate. The research pipeline encompasses exploratory data analysis, SMOTE-based class imbalance handling, feature engineering, MinMax normalization, and ANN model training with early stopping. Model performance is evaluated using accuracy, precision, recall, F1-score, and confusion matrix visualization. To address transparency limitations of black-box models, Explainable AI (XAI) techniques — SHAP and LIME — are incorporated, identifying balance differences, transaction amounts, and transaction type as the key fraud indicators. The findings confirm that ML-based approaches substantially outperform traditional rule-based systems, and that explainability tools are essential for building stakeholder trust in financial AI systems.