ATRASS #9: On the Trustworthy AI Dimensions Fairness, Confidentiality and Transparency

On the Trustworthy AI Dimensions Fairness, Confidentiality and Transparency and Their Dependencies

By Cor Veenman, senior scientist responsible AI, TNO

At TNO we developed a roadmap on trustworthy adaptive AI for autonomous systems and decision support systems. We set up an extensive research program across the domains covered by our organization to take up the challenges in system adaptivity and the trustworthy AI dimensions: fairness, confidentiality, and transparency. The program involves collaborations with universities for academic
research, applied research in Flagship projects, and use-case oriented research in use-case projects. In this talk, I will present results on this depth and breadth of research including the interplay between the trustworthiness dimensions in the scope of decision support systems. Examples are the development of an algorithm for the generation of models that mitigate (intersectional) fairness, to support fair model selection through visualization, to support fair data representation through visualization, and for counterfactual explanations that supports privacy of the counterfactuals. As the AI Act enforces high risk AI systems to implement the trustworthiness dimensions this research becomes increasingly urgent. Especially, for specific use case contexts there is the need for methods that focus on the tension between the trustworthiness
dimensions, which is the goal in the last phase of our roadmap."