In today's rapidly evolving technological landscape, the financial industry has embraced various advanced online services, including credit cards, ATM cards, online payments, and transfers. However, with these advancements, the risks and threats have also increased. This research thesis aims to find ways to protect customer accounts from hackers who exploit loopholes to steal account information. We have developed a machine learning model for anomaly detection in each account transaction to filter suspicious activities before confirming the transaction. When the model detects a risky transaction, the system alerts the account owner to recognize and verify the transaction before proceeding. Our research involved training models using a dataset of more than 12 million transactions and 12 features to apply the Isolation Forest (IF), Local Outlier Factor (LOF), and Autoencoder (AE) models to train. The results indicate that the AE model outperformed the others, with a Receiver Operating Characteristic (ROC) score of 0.92, compared to IF's 0.54 and LOF's 0.50. Training times varied, with IF taking 3.37 minutes, LOF 17.41 minutes, and AE 8.57 minutes. We selected the AE model as the most ROC for anomaly detection based on these findings. Subsequently, we developed an Application Programming Interface (API) based on the AE model to enhance banking systems' security, trust, and efficiency.