Detection of Credit Card Fraud in a Brazilian database using Autoencoder Neural Network

Abstract

The increasing number of credit-card transactions made over the internet in recent years has lead to a rise in the same proportion in the amount of fraud. Due to the large volume of web-based transactions that should be carried out daily, it is necessary to have a robust system to predict such crime to reduce loss and increase the confidence of banks and issuers. Deep Learning techniques emerge as a way to automate this process, training classifiers with data from past transactions to try to predict future frauds. In this paper, we build an Autoencoder model and perform a threshold tuning to predict fraudulent transactions. A proprietary Brazilian credit-card transaction database was used for training and performance evaluation of the model, containing almost 40 million transactions and challenging frauds, which were not previously detected by the organization's current fraud detection systems. The results of the experiments presented satisfactory results in a real-world dataset. The Autoencoder metrics show a good performance in fraud classification, reaching a positive Matthews Correlation Coefficient value and an AUC of 0.81, which are not affected by the database imbalance.