Enhanced Deep Learning Models for Secure and Efficient Cross-Border Financial Transactions

Abstract

As reliance on foreign financial transactions keeps increasing, the number of problems regarding security vulnerabilities and the identification of dishonest activity has expanded. Traditional fraud detection systems’ incapacity to identify fraudulent activity in real time and to adapt with the times may lead to severe financial losses. Including modern deep learning methods will help to offer a possible solution that will improve the accuracy and speed of the fraud transaction identification. This work attempts to provide a special hybrid model combining Genetic Algorithm-based feature selection with modified loss function (EL-UXGB) with Extreme Gradient Boosting (XGBoost). Deep Belief Networks (DBN), which handle the input features gathered from Internet of Things (IoT) enabled devices, also serve to improve detection skills. The 20,000 financial transactions used for model testing included 10% of which were deemed to be fraudulent. With a detection accuracy of 99.4%, a precision of 98.7%, and an F1-score of 98.9%, the proposed approach outperformed proven approaches including Logistic Regression (85.3% of the time) and Random Forest (91.6% of the time). Thirty percent drop in processing latency helped to demonstrate the real-time usefulness of the model. This enabled fast fraud detection free of scalability compromise. The results guarantee the security and efficiency of worldwide transactions by helping to clarify the capabilities of deep learning approaches to battle the growing complexity of bank fraud systems. Future research will largely concentrate on optimizing further and using it in the actual world over global financial systems.