Robust AI Techniques to Support High-consequence Applications in the Cyber Age

Abstract

Many of today's high-consequence tasks require narrow subject matter expertise (SME), tooling, and thoughtful thorough planning to transform intricate calculations and analysis into subsequent plans and actions. For AI to help tackle these types of problems, we must employ measures that ensure statistical and mathematical robustness every step of the way. This chapter will outline the motivations behind developing robustness and explainability measures for deep learning models and why they are important in the space of high-consequence ML. It will then overview the latest in the field of both explainability and confidence measures to quantify model robustness using a case study in nonintrusive load monitoring (NILM) for smart-grid control. We will then dive into how SME can be utilized to build and fine-tune robust foundation models with specialized vocabularies and modalities, along with concrete case studies of doing so in the field of nuclear nonproliferation and aviation. Finally, we will wrap up with vulnerabilities these foundation-model-forward approaches incur, how they may be attacked or undermined, along with what mitigations should be considered to prevent their circumvention.