2019 • Conference Paper
Compressive k-Means with Differential Privacy
Authors:
Schellekens, Vincent,
Chatalic, Antoine,
Houssiau, Florimont,
de Montjoye, Yves-Alexandre,
Jacques, Laurent ,
Gribonval, Rémi
Published in:
"Proceedings of SPARS'19"
Volume: 1 • Number: 1 • Pages: 1
In the compressive learning framework, one harshly compresses a whole training dataset into a single vector of generalized random moments, the sketch, from which a learning task can subsequently be performed. We prove that this loss of information can be leveraged to design a differentially private mechanism, and study empirically the privacy-utility tradeoff for the k-means clustering problem.