Regularized Loss Minimizers with Local Data Obfuscation
share
While data privacy has been studied for more than a decade, it is still unclear how data privacy will counter data utility. We focus on the balance of data privacy and data utility in this paper. We show that there are several regularized loss minimization problems that can use locally perturbed data with theoretical guarantees for loss consistency. Our result quantitatively connects the convergence rate of the learning problems to the impossibility of any adversary for recovering the original data from perturbed observations.
READ FULL TEXT