A semi-decentralised framework to enhance credit card fraud detection in a privacy-preserving manner using machine learning approach

Abstract

In today’s world, financial institutions rely massively on Internet banking to efficiently make services available to customers. Due to the increase in credit card use, the number of frauds has also escalated. Research has been done using state-of-the-art techniques to effectively avoid credit card fraud. The proportion of fraudulent and legit transaction in financial (credit card) data is unbalanced, and Data produced per second from the banking system are substantial in number, diversified, and heterogeneous. Therefore, the conventional methods of fraud detection have been proven to be challenging to tackle all the challenges. Adapting machine learning techniques to credit card fraud detection reduces many obstacles, as illustrated in the literature. Although machine learning techniques are efficient in credit card fraud detection, there are some additional challenges researchers must face, the imbalanced data is one of them. The researchers proposed collecting large amounts of data to balance the data distribution. One of the ways to address the issue of imbalance is to share data from different institutions or organisations. But due to Different regulation, the challenge is the data sharing by banks and financial institutes to access to the continuous stream of real-time data to learn the new heterogeneous patterns. The research was conducted to understand how data can be collect from different institutions without breaking different regulations like GDPR, or Data Protection Act 2018 in a privacy-preserving manner. This study introduces a framework that merges federated learning with other machine learning algorithms to improve credit card fraud detection, while also maintaining the privacy of data across the collaborating institutions. In the research analysis, different machine learning models such as random forest (RF), logistic regression (LR), multilayer perceptron (MLP), artificial neural networks (ANN), support vector machine (SVM), and decision tree (DT), were used. Along with these classical models, the research also experimented with the combination of Federated learning framework. The research also analysed the model's performance with and without the combination where it’s been found that ANN, random forest and XGboost performed well. The contribution of the research is a combined approach to integrating the machine learning model with the decentralised approach of the federated learning framework. In this approach, data are trained locally and only the trained model from the individual device is combined on a central server while keeping data privacy. The secure aggregation encryption method used by the FL framework ensures the end-to-end model transmission privacy between the clients and the server, thus following GDPR. The future scope of the contributed framework will allow banks and financial institutes to share the data collaboratively to train the machine learning model for effective CCFD. The initial experiment and results show that the performance of the artificial neural network, random forest and XGboost performed well in combination with federated learning provides better results than the other machine learning algorithms.