Enhancing Banking Security: A Machine Learning Approach to Fraud Detection

Abstract

Detecting fraudulent transactions in the cybersecurity industry is crucial, especially in banking, where legitimate transactions vastly outnumber fraudulent ones, creating highly imbalanced datasets. This study compares the effectiveness of two approaches, random under-sampling and Synthetic Minority Over-sampling Technique, to evaluate the accuracy of fraud detection in such datasets. Random under-sampling excludes instances from the majority class, improving recall but compromising precision. Conversely, SMOTE generates synthetic examples for the minority class, yielding a balanced F1 score and improved overall accuracy but with slightly reduced recall. To further enhance fraud detection performance, multiple machine learning models were evaluated to balance recall, precision, and F1 scores effectively. This research provides insights into the nuances of under-sampling and oversampling in fraud detection, guiding cybersecurity professionals in selecting techniques that best suit their organizational needs. It underscores the importance of addressing class imbalances and emphasizes the role of ongoing model refinement in achieving reliable fraud detection.