The substantial research challenge of identifying rare yet critical fraudulent transactions within highly imbalanced datasets is addressed by employing machine learning to detect fraudulent activities in banking data. The objective of this investigation is to create a fraud detection model that is both precise and efficient, thereby reducing the number of false alarms in banking systems and minimizing financial losses. The proposed methodology begins with the preprocessing of data, the selection of features, and the rectification of class imbalances using SMOTE. Subsequently, it implements supervised machine learning algorithms, including Random Forest, Support Vector Machine, Logistic Regression, and XGBoost. The performance of a model is assessed using metrics such as precision, recall, F1- score, and ROC-AUC. Experimental results suggest that ensemble methods, particularly XGBoost and Random Forest, outperform conventional models in the detection of fraud and the attainment of accuracy, while simultaneously generating fewer false positives. This demonstrates their ability to identify complex fraud patterns by utilizing transactional data. This research is essential because it improves the safety and dependability of digital banking systems, facilitates the detection of fraud in real time, and helps financial institutions prevent monetary losses while preserving customer trust. The adaptability of the proposed system to emerging fraud schemes is believed to make it capable of functioning effectively in modern financial environments.