Governing the Tradeoff Between Predictive Accuracy and Adversarial Robustness: An Enterprise Model Risk Framework

Machine learning models used in financial decision systems create a measurable risk tradeoff between predictive performance and resilience to adversarial manipulation. This study presents a governance framework for managing that tradeoff through stepwise adversarial hardening guided by feature importance. It introduces three enterprise risk measures—Management Attack Success Risk (MASR), Robustness Overfitting Risk (ROR), and the Robustness Accuracy Exchange Score (RAES)—that quantify residual adversarial risk and the business cost of robustness controls. The proposed measures provide decision-oriented evidence for risk governance, risk acceptance, and control oversight.