Drift-aware and privacy-preserving explainable credit scoring for fair decision-making in real world deployment

Credit scoring systems are used in a real-world financial environment faces challenges like data drift, privacy restrictions, fairness, and explainability regulations. However, current existing approaches optimise these goals separately, which limits their usefulness in real-world deployment situations. This study proposes a unified framework which integrates drift-aware modeling, privacy-preserving learning, and explainable-AI for robust credit risk assessment. A centralized Stochastic Gradient Descent (SGD) classifier baseline model was trained on aggregated data from a publicly available credit dataset from Lending Club, followed by federated learning approaches using Federated Averaging (FedAvg) and Federated Proximal Optimization (FedProx) to enable collaborative training across distributed data silos without sharing raw data. Additionally, Differentially Private Stochastic Gradient Descent (DP-SGD) was employed to enforce formal privacy guarantees by injecting controlled noise during model optimization. Experiments on time-evolving credit datasets simulated distribution shifts across temporal and demographic segments. Performance was evaluated across four dimensions like predictive utility using Area Under the Receiver Operating Characteristic Curve (AUC) and Precision-Recall Area Under Curve (PR-AUC), fairness using equalized odds gap and subgroup calibration, privacy risk via membership inference attacks, and explanation quality in terms of faithfulness, stability, and actionability. The centralized SGD classifier achieved an AUC of 0.96, while federated learning models achieved comparable performance with improved fairness under drift. Differential privacy reduced privacy leakage at the cost of moderate performance degradation. Explanation stability decreased under severe distribution shifts. The results highlight Pareto trade-offs across objectives, emphasizing the need for deployment-specific model selection in responsible credit decision-making systems.