Financial Fraud Detection Model using Recurrent Neural Network

Abstract

Financial fraud continues to pose a serious concern; each year, millions of cases are recorded. Fraudsters still use strategies to maneuver current security solutions such as two-factor authentication. This study offers a strong fraud detection system that examines user behaviour and transaction patterns using Deep Learning algorithm -Recurrent Neural Networks (RNNs). RNNsbased model was trained using the "Synthetic Financial Datasets For Fraud Detection" obtainedfrom one of the world renown data repositories, Kaggle.com. The dataset contains 11 columns (features) and 6,362,620 rows. To improve data quality, preprocessing operations- cleaning and feature engineeringwere carried out. The dataset was split into training and testing sets using ratio 80:20. Early stopping and hyperparameter adjustment were used to maximize the model's performance and to avoid overfitting. Millions of artificial transaction records that replicate actual financial transactions were created using the PaySim simulator and included in the collection.The model achieved Accuracy of 99.96%, Receiver Operating Characteristic Area Under the Curve (ROC AUC) of 99.69%, and Precision of 99.53% on the testing set. These outcomes demonstrate how well the model detects fraudulent transactions. The new model has the ability to drastically cut down financial losses dueto fraud, as evidenced by its great accuracy and precision.