On the Use of a Sequential Deep Learning Scheme for Financial Fraud Detection

Forecasting fraud detection has never been more essential for the finance industry than today. The detection of fraud has been a major concern for the banking industry due to the high impact on banks' revenues and reputation. Fraud can be related with an augmented financial risk, which is often underestimated until it is too late. Recently, deep learning models have been introduced to detect and forecast possible fraud transactions with increased efficiency compared to the conventional machine learning methods and statistics. Such methods gain significant popularity due to their ability to estimate the unknown distribution of the collected data, thus, increasing their capability of detecting more complex fraud events. In this paper, we introduce a novel multistage deep learning model that combines a feature selection process upon an Autoencoder model and a deep convolutional neural network to detect frauds. To manage highly unbalanced datasets, we rely on the Synthetic Minority Over-sampling Technique (SMOTE) to oversample our dataset and adjust the class distribution delivering an efficient classification approach. We describe the problem under consideration and our contribution that provides a solution for it. An extensive set of experimental scenarios are adopted to reveal the performance of the proposed scheme exposing the relevant numerical results. A comparative assessment is used for proving the superiority of our model compared with a Support Vector Machine (SVM) scheme, a classical CNN model and the results of two researches that use the same dataset.