Bank Credit Default Risk Assessment Model Based on Federated Learning

This paper presents a privacy-compliant credit default risk model employing Vertical Federated Learning (VFL), XGBoost, Knowledge Distillation (KD), and Temperature Scaling (TS) techniques. The aim is to address the challenge of maintaining privacy while improving model performance in federated learning environments, particularly focusing on bank credit risk prediction. The model is tested on the CRMS-2024 dataset. Compared to local models and existing federated learning methods, the proposed model shows significant improvements across multiple metrics. The proposed method achieves an AUC of 0.804, 0.063 higher than the local model, and reduces the expected calibration error (ECE) to 0.024. The model also demonstrates excellent fairness performance. With temperature scaling, the demographic equilibrium difference (DPD) is 0.021, and the chance equality gap is 0.028. Statistical significance is evaluated using the DeLong test for AUC and the BCa bootstrap method for Brier scores, and Holm correction is applied for multiple comparisons. The proposed method remains robust to noise, data imbalance, and Byzantine attacks, demonstrating the performance and calibration improvements of KD and TS in non-IID federated environments. This paper also explores future improvement paths, including cross-domain validation for datasets such as LendingClub, and integrates regulatory APIs to achieve compliance.