Verification, Validation, Assurance, and Trust of Machine Learning Models and Data for Safety-Critical Applications in Armaments Systems

To address the challenges posed by verification and validation (V&V) of artificial-intelligence (AI) and machine-learning (ML) for high assurance in safety-critical functions of armaments systems, ARiA developed and demonstrated the feasibility of the Machine Learning Qualification Process (MLQP). Based on the promising findings and technical feasibility demonstrated during Phase I, ARiA proposes to further develop MLQP during Phase II along with means and methods to evaluate the metrics and measures developed during Phase I. The Phase II effort will result in documented processes and procedures for V&V of AI/ML systems and their components—including data, features, and models—when used in safety-critical functions of armaments systems. At the conclusion of Phase II the MLQP processes and procedures will be given a pilot demonstration using prototype implementations of artifact templates and software tooling for evaluating metrics and measures. MLQP development will be accompanied by creation of formal documentation and guidance for applying MLQP throughout the acquisition lifecycle (including design during initial S&T funding and development during later RDT&E funding) as well as over the lifetime of a fielded system (during O&M funding to guide periodic recertification). Building on this guidance, the pilot demonstration will showcase use of MLQP throughout the lifecycle of an AI/ML system using a range of relevant data spanning multiple sensor modalities and data formats (images, time series, tabular, etc.).