E-banking has significantly enhanced customer satisfaction by providing improved service quality and convenience. For banks, it has become a crucial tool in gaining a competitive edge in the industry. However, the rise of e-banking has also attracted the attention of fraudsters, leading to growing concerns about security. Many potential users remain hesitant to adopt e-banking services due to the perceived lack of adequate security measures. Despite these challenges, e-banking continues to thrive. This paper delves into the security issues surrounding e-banking, exploring both the vulnerabilities and the methods used by cybercriminals. Furthermore, it examines the common challenges associated with e-banking fraud and highlights the essential features that must be addressed to combat these threats effectively. Consumers must verify that their payments are directed solely to the intended service provider to avoid online fraud, which can lead to data compromise and the inconvenience of reporting and blocking payments. In our approach, we apply Bayesian optimization to refine hyper parameters, taking into account real-world challenges such as handling imbalanced datasets. This method ensures a more efficient and targeted search for optimal hyper parameters, addressing key practical concerns that may arise during model training. We suggest weight-tuning as a pre-processing step for unbalanced data and integrate CatBoost and XGBoost to increase the efficiency of the LightGBM method by incorporating a voting process. Additionally, we employ deep learning to further selecting hyperparameters, specifically focusing on our proposed weight-tuning technique. Experimental validation is conducted on real-world datasets, where we employ recall-precision metrics along with the standard ROC-AUC to better address unbalanced datasets. CatBoost, LightGBM, and XGBoost are individually evaluated using a fivefold cross-validation approach.