Federated Learning-Based Credit Card Fraud Detection: Performance Analysis with Sampling Methods and Deep Learning Algorithms

• Published in: 2023 IEEE International Conference on Cyber Security and Resilience (CSR) pp. 180 - 186
• Main Authors: Aurna, Nahid Ferdous, Hossain, Md Delwar, Taenaka, Yuzo, Kadobayashi, Youki
• Format: Conference Proceeding
• Language: English
• Published: IEEE 31.07.2023
• Subjects: CNN; Credit card fraud; Credit cards; Data models; Data privacy; Deep learning; Federated learning; fraud detection system; LSTM; MLP; Sampling methods; Turning
• DOI: 10.1109/CSR57506.2023.10224978

The exponential technological advancement is turning everyone towards an easy and efficient way of financial transactions. Consequently, the use of credit cards is rising substantively, creating a more incredible opportunity for fraudsters which is an alarming concern nowadays since a fraudster may use several tools, techniques and tactics to make a fraudulent transaction. As a countermeasure, an effective fraud detection mechanism and highly sensitive data privacy preservation are imperative to detect fraudulent transactions. This paper proposes a Federated Learning (FL)-based fraud detection system since its key feature preserves the privacy of highly sensitive data, wherein the model could be trained without sharing the credit card data in the cloud. We contemplate three Deep Learning (DL) models: Convolutional Neural Network (CNN), Multi-Layer Perceptron (MLP), and Long Short-Term Memory (LSTM) regarding the FL approach. Subsequently, to overcome the data imbalance issue, four distinct sampling techniques are explored to inspect the impact on the traditional centralized and FL approaches. Finally, we further investigate and compare FL-based detection systems with diversified state-of-the-art models. Our experimental results demonstrate that the proposed method is superior compared with state-of-the-art methods and achieves high detection rate of 99.51%, 98.77% and 98.20% respectively for CNN, MLP and LSTM models.