One of the currently visible requirement of data analysis is to cope with data privacy requirements. The quest for privacy triggers design challenges in the construction of Machine Learning models. In this study, we propose a novel and systematic design methodology of fuzzy rule-based models completed in the following privacy-driven environment. Multidimensional data are available for model design in two different formats: those attributes that are privacy sensitive are provided at the higher non-numeric level of abstraction represented in the form of information granules while the remaining ones are available directly. It is shown that the granular manifestation of data becomes beneficial in the formation of the conditions of the rules whereas numeric attributes are available for the construction of the conclusions of the rules. The generic development scheme is discussed. Some generalization is introduced to cope with a number of data islands associated with corresponding data owners. It is shown that a formation of the data structure calls for a mechanism of collaborative clustering. A way of handling information granules at different levels of information granularity is discussed. Some illustrative examples are presented.