Fraud Detection Solution for Monetary Transactions with Autoencoders

Abstract

Fraud has turned into a trillion-dollar industry which may lead to risk of financial loss as well as the loss of customers' and stakeholders' confidence on financial organizations. Nowadays, online transactions, mobile wallets and payment card transactions are becoming more popular within society. With the growth of such cashless transactions, the number of fraudulent activities in the world is also increasing. According to the current global economic context, efforts being made to detect and prevent frauds are also increasing. Having an effective financial transaction fraud detection system could save trillions of dollars from fraudulent activities. Supervised machine learning based fraud detection solution is the trending mechanism used in fraud detection solutions. Nevertheless, such supervised machine learning based solutions need a labelled dataset in order to train the machine learning model. The reason for the existence of current fraudulent actions is that labelled datasets are hard to find in real-world environments, and if such labelled datasets are available, thereafter such fraud detection solutions would detect fraudulent patterns based on the fraudulent patterns of the fraudulent events in the training labelled dataset. Therefore, there is an extensive business requirement of having a fraud detection solution which can be trained using a raw financial transaction dataset, in other words using an unlabelled dataset which is commonly available in financial transaction systems in order to detect accurate fraudulent events. Test results obtained for the synthetically generated dataset shows that autoencoder is able to detect fraudulent transaction events with 83% of AUC score which represents high capability of binary classification as fraudulent or genuine transactions.