1: \begin{abstract}

2:     We study statistical risk minimization problems under a privacy

3:     model in which the data is kept confidential even from the learner.

4:     In this local privacy framework, we establish sharp upper and lower

5:     bounds on the convergence rates of statistical estimation

6:     procedures.  As a consequence, we exhibit a precise tradeoff between

7:     the amount of privacy the data preserves and the utility, as

8:     measured by convergence rate, of any statistical estimator or

9:     learning procedure.

10:   \end{abstract}

11: