PRIVACY-PRESERVING EXPLAINABLE DEEP LEARNING FRAMEWORK FOR LOAN DEFAULT RISK PREDICTION

Abstract

Automated loan-approval systems increasingly rely on deep neural networks that can capture non-linear relationships between an applicant's financial history and their likelihood of default. Two obligations compete inside such a system: the model must be accurate enough to be economically useful, and it must not leak information about the individual applicants whose records it was trained on. Differential privacy (DP) is the standard mathematical language for the second obligation, but in the bank-loan-prediction literature it has almost always been bolted onto the outside of a classical ensemble model  noise is added to features or to a tree ensemble's outputs  rather than integrated into the training procedure of a deep network itself. This study closes that gap by training a multilayer perceptron (MLP) for loan default prediction with DP-SGD, the differentially private stochastic-gradient-descent procedure in which every individual applicant's gradient contribution is clipped and Gaussian noise is added before the network's weights are updated, so that the privacy guarantee is a property of the learning algorithm rather than of a pre-processing step. On the standard 614-record public loan-approval benchmark, a non-private baseline MLP reaches 83.3% accuracy and a ROC-AUC of 0.775 (mean of five seeds), statistically indistinguishable from a Random Forest (85.0%) and a histogram gradient-boosting model (81.3%) trained on the identical split. Training the same architecture with DP-SGD across a swept noise level traces a smooth, monotonically increasing privacy-utility curve: accuracy rises from 48.5% at a strict privacy budget of ε = 0.39 to 82.0% at a more permissive ε = 7.27, recovering 98% of the non-private model's accuracy. Because the advanced-composition bound used for that headline sweep is deliberately conservative, a second, Rényi-DP accountant is computed for the same trained networks and reports privacy budgets four to five times tighter for identical noise levels, and the qualitative collapse-then-recover pattern replicates on an independent 572-record external validation dataset. Three complementary explainability techniques  permutation importance, integrated gradients and SHAP  are applied directly to the trained network rather than to a separate surrogate tree model, and all three independently identify credit history as the dominant predictor. A secondary experiment compares three ways of handling the dataset's 2.2:1 class imbalance and finds that synthetic oversampling and class-weighting trade a small amount of aggregate accuracy for a modest improvement in the minority-class recall. The results suggest that DP-SGD is a practical, feature-agnostic route to privacy-preserving deep learning for credit-risk prediction.