A Proposed Approach for Certifying Machine Learning Pipelines Containing Multiple Trust Objectives

We propose to develop methods for certifying Machine Learning (ML) models by optimizing multiple trust and decision objectives. By combining measures of trustworthiness with domain- specific decision objectives, we aim to establish the criteria necessary to meet exacting standards for high-consequence applications of ML in support of national security priorities. In accordance with Executive Order-14110, our objective is to promote the safe, secure, and trustworthy development and use of Artificial Intelligence (AI) by delivering a generalizable process that can readily inform ML certification. Current credibility assessments for ML are developer-focused, application-specific, and limited to uncertainty quantification (UQ) and validation, which are necessary but insufficient for certification. Whereas ML developers are primarily concerned with various measurements of model accuracy, non-ML- expert stakeholders are concerned with real-world quantities such as risk, or safety; this suggests that a more holistic technical basis is needed. With multiple objectives, decisions may only be Pareto-optimal: no objective can be improved without making another worse. Designing certification processes with multi-objective design optimization allows for the balancing of requirements for specific applications. The absence of evidence-based, generalizable approaches to certification restricts our ability to develop and deploy credible, mature ML models. To address this gap, we will operationalize AI risk-frameworks, such as the National Institute of Standards and Technology’s Risk Management Framework, by synthesizing important Trustworthy AI capabilities such as robustness testing and anomaly detection. To validate and refine our approach, we will engage in collaborative case studies applying our tools to real-world datasets and scenarios while gathering feedback to ensure their effectiveness and usability. Our primary research goal is to develop a process for designing certifications for trustworthiness of ML, particularly in high-consequence applications. Drawing inspiration from the multi-tiered approach of software testing, which encompasses unit to integration tests, our strategy involves assessing trustworthiness throughout the ML development lifecycle and conducting system-level, evaluations of non-functional properties such as safety, fairness, and privacy. If successful, our research will position ML to fulfill the requirements of high- consequence domains, as evidenced by a measurable improvement in the reliability properties of our exemplar models.