Detecting Fraud in Automated Teller Machine Transactions in the Nigerian Bank System Using Unsupervised Deep Learning

Automated Teller Machines (ATMs) are globally adopted for financial transactions. However, the exponential rise in ATM fraud poses a significant risk to Nigeria's digital payment and banking systems, with many banks failing to provide intelligent services to ATM customers. Moreover, existing research predominantly employed supervised learning on labelled data for ATM fraud detection, overlooking unlabelled customer-clustered transaction data. In this study, customer-clustered transaction attributes were employed with unsupervised autoencoder deep learning models to develop a fraud detection system. Three years of historical transaction data were obtained from a selected Nigerian bank's ATM transaction repository, consisting of 1.2 million transaction records from 10 bank customers. The transaction features considered in this work included CardNo, Amount, TransType, BranchCode, and TransDateTime. H2O autoencoder deep learning models, and anomaly detection algorithms were adopted to uncover underlying patterns in ATM transactions. The results indicated a remarkable overall system performance, with an average Mean Squared Error (MSE) of 0.0057, accuracy of 97.6%, precision of 94.8%, recall of 93.5%, and an F1-score of 94.0%. Individual model assessments revealed slightly lower performance for specific customer clusters. The study's findings emphasized the efficacy of unsupervised deep learning in combatting ATM fraud within the Nigerian banking system. Keywords: ATM, H2O.ai, Unsupervised Learning, Deep Learning, Autoencoder, Fraud detection CISDI Journal Reference Format Ojulari, H.O., Oke, A.O. & Arulogun, O.T. (2024): Detecting Fraud in Automated Teller Machine Transactions in the Nigerian Bank System Using Unsupervised Deep Learning. Computing, Information Systems, Development Informatics & Allied Research Journal. Vol 15 No 2, Pp 1-... dx.doi.org. Available online at isteams.net