Federated learning for credit card fraud detection: A privacy-preserving approach with controlled noise integration

With the rapid increase in e-payment technology, cards are present as one of the essential tools. However, with this growth comes a risk of fraudulent attacks that may cause involved parties' losses and damage. Banks aim to establish strong fraud detection systems to protect assets and obey regulator rules. Therefore, developing a model that copes with security and integrity requirements is crucial. This research project introduced advanced Machine Learning (ML) models, such as Deep Neural Network (DNN) and ensemble learning with AdaBoost, to detect fraud while addressing the skewed nature of datasets using balancing methods. Additionally, it facilitates collaborative learning among banks using Federated Learning (FL) while preserving data privacy. In this study, the FL model was tested against various percentages of label-flip attacks to evaluate resilience against malicious acts by banks trying to sabotage the learning process. The models were tested on two datasets: a real dataset from an Iraqi bank and the known Kaggle creditcard dataset. Models were assessed on a set of performance metrics to cover all aspects of the methods. Results showed that ensemble learning with Random Forest (RF) and AdaBoost achieved remarkable performance across both datasets. Moreover, the FL cosine-based worked better than the existing Federated Average method. Lastly, the proposed approach combining RF+AdaBoost with FL cosine aggregation surpassed the existing method when validated on a private bank dataset achieving 96.47% accuracy and 94.58% recall. The present study contributed to the academic literature by addressing the lack of real datasets in the fraud detection domain.