Enhancing Credit Risk Assessment with Federated Learning Through a Comparative Study

Digital transformation in the financial sector heightens the need for privacy-preserving methods in credit risk assessment. In this chapter, we proposed a federated-learning-based risk prediction model to predict credit card defaults, while preserving customer data privacy. We utilized “Default of Credit Card Clients in Taiwan” dataset and applied preprocessing techniques. We performed sampling methods to handle the class imbalance issue. Also, we employed feature reduction methods to optimize model performance. To evaluate the Central ML, FedAvg, and FedF1 methods, we built and compared five machine learning algorithms, such as logistic regression, multilayer perceptron, support vector machine, XGBoost, and random forest. Our findings show that FL approaches can maintain competitive performance compared to Central ML methods, while preserving data privacy and can utilize more data from different clients. Also, the novel FedF1 method provides comparable results to central ML models. Project code is publicly available at: https://github.com/Mstfakts/Federated-Learning-Comparative-Study.