Smart Transactional Fraud Detection using Artificial Intelligence Techniques

Abstract

The banking industry significantly contributes to a nation s development and eco newlinenomic improvement, but the rise of smart transactions in recent times where customers newlineconduct trade both online and physically, has raised concerns about fraud. The availabil- newlineity of advanced technology has made online banking a prominent channel for business newlinebut it has also resulted in a surge of fake banking activities and fraudulent transactions compromising the security and trust of users. These fraudulent activities such as hacking user information, spreading viruses and scams, and creating fake links have resulted in significant losses. newlineTo address the challenges of Fraud Detection (FD) in smart payment transactions, newlinethis study proposes four contributions. The first contribution introduces the Refitted newlineAdaptive Synthetic Algorithm (RAA) to tackle data imbalance problems. This modified newlineAdaptive Synthetic Oversampling (AdaSyn) algorithm leverages the Gaussian distribu- newlinetion to generate mock data that aligns better with the minority samples distribution. newlineSkewed class distributions often bias the results of Machine Learning (ML) and Deep newlineLearning (DL) algorithms that typically use overall accuracy as a performance index. newlineRAA outperforms other under and oversampling methods and effectively generates syn- newlinethetic data without noise or bridges. newlineThe second contribution involves the creation of a novel model called RDQN (Re- newlineinforcement Learning (RL) with Rough Set Theory(RST)). This model combines RST newlineand deep reinforcement learning to enhance fraud detection. The model employs three newlinesteps: data pre-processing using RST to enhance data quality, combining DNN and Q newlinelearning to create a Deep Q-Network. This approach accurately classifies and predicts newlinetransaction categories by activating the reward function through agents making it difficult for fraudsters to escape from detection. The model achieves a remarkable accuracy of 96.09% and by incorporating feature selection, it effectively minimizes processing time as well. newlineThe third cont