Trustworthy AI - Managing Privacy and Explainability in Collaborative Data Analysis

This talk will present the paradigm of privacy preserving collaborative data analysis, discussing how to measure privacy gain and data utility in those problems where it is necessary to put together data, in a collaborative fashion, among different stakeholders, without base trust assumptions and where it is not possible to exploit federated learning approaches. The talk will present an approach to trade-off the two measures of data privacy and data utility exploiting the compatibility matrix concept. Then we will introduce the concept of explainability as cornerstone of the trustworthy AI, and the effect of mechanisms to enforce explainability, on data utility and data privacy, explaining how to include also this dimension in the trade-off problem.