Law enforcement authorities continually confront new and vexatious frauds involving computer virus attacks or credit card frauds committed against financial institutions and banking card companies. Such phenomenon used to commit online banking fraudulent transactions are occurring continuously and lead to a loss of gargantuan amount of money. Information security specialists and associations usually refer to the process of selecting practitioners as safeguards of fraud assessment, fraud analysis, or fraud detection. Designating such safeguards based on fraud calculations can he a exasperating and high-price process over the past few decades. This paper aims to understand how deep learning (DL) models can be benevolent in detecting fraudulent transactions with high accuracy. Dataset is extracted from one month of genuine financial logs of a mobile money service company in Africa, containing about more than six million transactions. The most satisfactory machine learning techniques as ensemble of decision tree (EDT), and deep learning techniques as stacked auto-encoders (SAE) and Restricted Boltzmann Machines (RBM) classifiers are applied on the preprocessed data. The performance of generated classifier models is evaluated based on accuracy, sensitivity, specificity, precision, confusion matrix and ROC values. The results of optimal accuracy are 90.49%, 80.52% and 91.53% respectively. The comparative results manifest that restricted Boltzmann machine performs superior than the other techniques.