Evaluation of machine learning classification models for detecting electronic fund transfers scam SMSes
Loading...
Date
2020
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
University of Namibia
Abstract
The last decade saw the emergence of mobile banking and a pervasive transcendence of spams from email to SMS communications. M-banking offers the users an ability to execute EFT transactions using mobile devices and allow them to receive SMS notifications acknowledging their transactions. While this provide convenience to m-banking users, in the wake of SMS spams it also presented vulnerabilities that could be exploited to scam money and goods from them. To execute these scams, spammers send forged EFT ( e.g. e-wallet) deposit notification SMSes to unsuspecting users, then contact and request them to do EFT payments as refunds for the supposed en-oneous deposits acknowledged by the bogus notifications. Similarly, during goods exchange, scammers use forged deposit notification SMSes to trick sellers to believe that they have paid for the goods. In Namibia, the high affordability of SIM cards and the readily available access tom-banking accounts such as e-wallet by anyone with a valid SIM number provides a favourab le operating environment for the EFT SMS scammers. Inferences from literatures on novel spam filtering techniques suggested that implementing machine learning classification could help address the EFT SMS scams problem, partly motivating this study to evaluate such application. Prevalent reporting of EFT SMS scams in local media (which mostly involves the country"s largest bank by
market share, FNB) and the observed lack of dedicated IT solutions to address such
problem were other factors that inspired this work. The study collected a dataset of
ham and EFT scam SMSes, from which machine learning features for classifying
SMSes were extracted. This was followed by a pre-evaluation to determine the features that allow ham and EFT scam SMSes to be classified optimally. SMSes comprising the collected dataset were then represented using the optimal features and used to train and evaluate Suppo1i Vector Machine, Naive Bayes and Random Forest classifiers. The evaluation results revealed that the SVM classifier was the most effective with respect to detecting EFT scam SMSes, achieving a FNR=0.00, CA=0.992, Recall= l .0 and F l-measure=0.995. The RF classifier followed with FNR=0.0 11 , CA=0.983, Recall=0.989 and F l -measure=0.989; while the NB classifier came last with FNR=0.027, CA=0.975, Recall=0.973 and F l -measure=0.983. The envisaged future work will look to use the methods, findings and conclusions drawn in this study to guide development of mobile application(s) that implement machine learning classification to detect EFT scam SMSes.
Description
A thesis submitted in partial fulfilment of the requirements for the degree of Master of Science (Information Technology)
Keywords
Mobile banking, Pervasive transcendence, EFT payments