Enhanced NBA Fraud Detection in Online Banking using Ensemble Learning and SMOTE

Online banking fraud, particularly New Bank Account (NBA) fraud, has led to substantial financial losses due to the rising adoption of digital banking. The imbalanced nature of NBA fraud datasets, where fraudulent cases are significantly fewer than non-fraudulent ones, poses challenges for machine learning models, often resulting in misclassification and reduced fraud detection rates. To mitigate this issue, SMOTE oversampling is applied to balance the dataset. Several Machine learning algorithms, including Binary Logistic Regression, K-Nearest Neighbors (KNN), Naive Bayes, Stacking Classifier (combining LR, NB, and KNN), and Voting Classifier (ensemble of Bagging Random Forest and Boosted Decision Tree), are explored to enhance detection accuracy. While the KNN model achieves 96% accuracy, ensemble methods improve performance, with the Voting Classifier reaching 99.1% accuracy by leveraging the combined strengths of multiple classifiers. The implementation of a Flask-based front-end with user authentication further facilitates real-time fraud detection and user testing. Integrating advanced ensemble techniques ensures a more robust and accurate fraud detection system, improving financial security and reducing fraudulent transactions in online banking.