Methodological Guideline for Assessing the Trustworthiness of Models of Artificial Intelligence with CAISAR

Models of Artificial Intelligence have found application in many areas, from perception to natural language
processing. However, their adoption in safety-critical applications require these models to exhibit properties
of trustworthiness. Motivated by decade-long advances in software reliability, formal methods have surged
as noteworthy approaches to assess such properties. This deliverable identifies three categories of properties
for trustworthy AI models, and provides a methodological guideline for their specification and assessment
by means of CAISAR, a software platform that federates multiple tools, based on formal methods, for the
trustworthiness of AI models.