Methodologies to Facilitate AI/ML Assurance for Safety-Critical Systems

Author(s):

James Michaelis

Abstract:

In line with U.S. Army research and modernization priorities, AI/ML systems have been acknowledged as imperative to enhance operational effectiveness and maintain strategic advantages on the battlefield. Since these systems commonly operate under missioncritical settings, methods to facilitate trust and risk assessment represent an essential Army requirement. To address such needs, ARL initiated an investigation into methodologies for definition and usage of safety argumentation over U.S. Army AI/ML technologies. FY25 program efforts have included the following activities: 1) investigating approaches for designing argument structures to establish claims of AI/ML safety backed up by corresponding evidence gathered through test, evaluation, validation, and verification (TEVV) activities; 2) defining approaches to enable encoding of assurance cases to facilitate their review by Army stakeholders and corresponding usage by AI/ML TEVV support services; and 3) conducting a set of engagements with research teams at ARL to obtain information on AI/ML technologies under development, which are intended to support assessment of the assurance case methodologies for the Army’s S&T needs.