Framework for Artificial Intelligence (AI) Resilience Through Evaluation of Systems and Technology (FAIREST)

The Framework for Artificial Intelligence (AI) Resilience Through Evaluation of Systems and Technology (FAIREST) aims to enable test planning of AI-enabled systems (AIES) for operational resilience and cyber survivability. Given the unique cyber risks and paradigms introduced by AIES (e.g. data drift/poisoning, adversarial manipulation, opaque decision-making), test designs for AIES resilience will require an evolved approach over traditional cyber-physical complex systems. AIES resilience engineering and testing will require new methodologies, metrics, and supporting environments, especially when considering cyber threats and viewing test as a continuum. Utilizing the Framework for Operational Resilience in Engineering and System Test (FOREST) [1] as a launch point, FAIREST utilizes systems-engineering-based approaches to define and demonstrate a methodology for AIES resilience test planning. This work first frames and scopes AI resilience concepts then defines and models the functional role of AI in resilience context. Subsequently, a taxonomy of AIES disturbances (i.e. “catalysts”) is established to better refine terminology and provide material suitable for design of experiments and/or risk assessment. The full FAIREST methodology is presented via nine meta-process elements called testable requirements elicitation elements in a non- linear, interconnected mapping spread across three testing phases: resilience planning, resilient response, and human-AI teaming. FAIREST is demonstrated on a simulated multi-agent minefield traversal under reduction in visibility. FAIREST offers contributions in domains including mission engineering and integration as well as cyber policy. Future research directions include focusing on where cyber and DevSecOps is relevant to AIES, exploring how to refine testable requirements, determining objective/threshold values for a model, and working directly with cyber test & evaluation stakeholders to model adversarial disturbances (i.e. cyberattacks) to further explore test planning design and candidate AI-based resilience mechanisms.