Abstract
The concept 'sensitivity' has multiple and sometimes incompatible usages and definitions, as they can be found in the scientific and technical literature. A strategy is proposed toward a conceptual framework in which sensitivity is qualitatively intended as a feature of a black box behavior and quantitatively is defined according to specific evaluation types (interval/ratio, ordinal, nominal) for both deterministic and stochastic behaviors. The proposed formal definitions characterize stochastic sensitivity as constituted of "effective" and "confounding" components, that can be simultaneously present and contribute to a desirable and unwanted increment of global sensitivity respectively. Two examples taken from the context of imaging systems and image-based measuring systems, in which sensitivity is computed in presence of non-negligible uncertainty sources, provide some hints on the usefulness of the proposed framework.