Speaker: Jan Ramon (Inria)

Title: Privacy preserving learning in the curious and not so honest setting

Abstract: In this presentation I will discuss privacy-preserving computation of statistics in environments where not all agents can be considered honest. In particular, I will review a number of existing directions (differential privacy, homomorphic computing, ...), highlight their shortcomings and propose new strategies in a search for an approach having all desired properties of privacy, security and accuracy. In that context, I'll also discuss advantages and disadvantages of centralization and coordination. This discussion then will lead to the GOPA algorithm which the Magnet team has been researching recently. I will conclude with some analysis and directions for future work.