Resilience Measurement Framework For Post-deployment Artificial Intelligence (ai) Integrated Systems

Abstract

Resilience is largely defined as the ability to adapt or recover from adverse conditions, stresses, attacks, or compromises on systems that use or are enabled by digital resources. In Artificial Intelligence Management and Research for Advanced Networked Testbed Hub (AMARANTH), resilience is measured in the amount of time it took from the beginning of a testing period for the model to reach predictions outside of the original 95% confidence interval or using the Kullback-Leibler (KL) divergence theorem, the Population Stability Index (PSI), and traditional methods such as root mean squared error (RMSE) threshold. Artificial Intelligence (AI) model drift is of significant concern when deploying AI-integrated systems into critical and/or secure environments. Drift can impact resilience of the AI-integrated system post-deployment and requires consistent maintenance and upkeep to ensure the model is accurate and precise. To quantify model drift and predict the point when a model's drift becomes unacceptable, we describe using Kullback-Leibler (KL) divergence, Population Stability Index (PSI) and/or confidence interval width estimations to determine the point of failure and time to failure of a model post-deployment. Through simple code functions, the KL-divergence, PSI, confidence interval, and root mean squared (RMSE) point of failures can be used to derive when a model needs to be maintained as well as the impact of adversarial action through statistical means.