International Journal of Science, Engineering and TechnologyAuthors: Dr. Gundupagi Manjunath, Dr. N Chandan Prashad
Abstract: The increasing digitization of financial services has led to the transformation of payment systems through improved convenience but also increased vulnerability to fraudulent attacks. Rule-based approaches to detecting such threats, limited by their static nature and preconceived threshold limits, prove inadequate to combat evolving threat models. In this paper, we propose a new approach to fraud detection through artificial intelligence that utilizes ensemble machine learning and deep neural networks to create a zero-trust architecture security mesh. Our system utilizes Random Forest and XGBoost classifiers along with LSTM sequence aware networks and adaptive learning capability for addressing concept drift. The experiments conducted on our algorithm based on the PaySim synthetic transaction data set demonstrate significantly better fraud detection performance, with an accuracy of 99.96%, precision of 91.84%, and recall of 89.12%. Our approach lowers false positive rates by 21% compared to traditional gradient boosting baselines while sustaining sub-120ms inference time.
DOI: http://doi.org/10.5281/zenodo.21676072
