a501bebf79d570651ff601788ea9d16d-AuthorFeedback.pdf

In the local DP set-up, data owners add noise totheir data entries and send them toa10 thirdparty(themodelholder/datacollector). The overall system issaid tobedifferentially private ifthe12 view of the data collector does not change much if the data set is modified in just one entry. The relation between DP and our work is an interesting one. Locally-private DP offers the possibility of obtaining50 statistics and ML models trained over a data set distributed over many owners in a private way. WebelievethatSMCandDPare56 complementary solutions; bringing the promise of privacy-preserving ML to practice will require an association of57 theseandotherparadigms.