Federated learning approach for credit card fraud detection

Abstract

Nowadays, an increased trend of credit card transactions occupying a more and more substantial role over cash, is observed. This trend is followed by the increasing expansion and evolution of the ways that fraudulent transactions can be performed. The more the anti-fraud systems evolve, the more sophisticated the fraud attacks become. Hence, this fraud and anti-fraud competition is highly reflected in the literature, establishing the variety of systems and models addressing this issue as an inevitable reality. Although systems evolve, the number of models addressing the problem from the perspective of data privacy protection is not as high as expected. Only a limited number of scientific papers can be found in the literature that take into consideration the factor of the data privacy, when solving the fraud detection problem. With regards to that, there is an obvious deficiency in the literature concerning the Federated Learning approach in the context of fraud detection, even though it has been successfully implemented in other fields such as Health and industry 4.0, where distribution of the learning on the one hand and privacy on the other are of the essence. In this Dissertation, we are trying to address the credit card Fraud detection problem using the Federated Learning approach. Thus, this Dissertation aims to contribute towards that direction, by examining the benefits and the performance of existing Federated learning models such as FedAvg on fraud detection and also by introducing novel approaches such as FedRandom Forest. FedRandom Forest combines one of the best performing algorithms, Random Forest, with the Federated Learning architecture, illustrating so competitive results as the existing literature that addresses the issue using Random Forest or ANN.