Explaining to defend: SHAP-based defense mechanism against adversarial attacks in P2P credit risk

Peer-to-peer (P2P) lending platforms, much like traditional financial institutions, rely heavily on credit risk assessment models to estimate the likelihood of borrower default. With the increasing integration of Artificial Intelligence (AI), particularly Machine Learning (ML), these institutions have embraced neural networks to enhance predictive performance. However, this technological shift introduces a critical vulnerability: susceptibility to adversarial attacks, intentional perturbations in input data designed to mislead ML models. This paper evaluates how such adversarial manipulations affect the reliability of neural network-based credit risk models in P2P lending and proposes a defense framework to mitigate these risks. Specifically, we develop and test a dual defense strategy combining adversarial training with a novel SHAP-based model-switching mechanism. The adversarial training improves resilience under attack but slightly reduces accuracy on clean data, whereas the SHAP-based switch compensates for this loss by detecting and redirecting perturbed inputs to a robust classifier. Compared to adversarial training alone, the proposed SHAP-based defense improves ROC AUC by +0.0541 on clean data while maintaining identical robustness under attack, demonstrating its superior ability to balance predictive performance and security. This quantitative gain highlights the efficiency of our hybrid defense relative to traditional single-model approaches. Experimental results show that the proposed approach preserves predictive accuracy on raw inputs while maintaining robustness against adversarial perturbations. Overall, our findings highlight a practical pathway toward building interpretable and secure AI-driven credit scoring systems for digital lending platforms.