Authors: Vani Nagendra, Dr. Geevarathna
Abstract: The quick growth of digital bank service offers creates, in its turn, greater chances for committing financial fraud, which, in turn, is very dangerous for both users and banks. In this regard, this paper suggests a deep learning algorithm that is designed to detect financial fraud in the sphere of e-banking. Such system employs a combined approach based on the use of LSTM neural networks along with CNN and an attention module in order to analyze both sequential and spatial data. The training procedure of the system was carried out on a large transaction dataset with skewed distribution via using appropriate oversampling techniques and cost-sensitive learning approaches. As a result, the system proved to be highly efficient, scoring 99.42% in accuracy, 98.87% in precision, 97.63% in recall, and 98.24% in F1-score. These figures significantly exceed the results provided by Random Forest, XGBoost, and simple LSTM models. Moreover, the system is equipped with SHAP analysis capabilities.
DOI: http://doi.org/10.5281/zenodo.20671872
