English

Efficiency in local differential privacy

Statistics Theory 2024-03-08 v3 Statistics Theory

Abstract

We develop a theory of asymptotic efficiency in regular parametric models when data confidentiality is ensured by local differential privacy (LDP). Even though efficient parameter estimation is a classical and well-studied problem in mathematical statistics, it leads to several non-trivial obstacles that need to be tackled when dealing with the LDP case. Starting from a standard parametric model P=(Pθ)θΘ\mathcal P=(P_\theta)_{\theta\in\Theta}, ΘRp\Theta\subseteq\mathbb R^p, for the iid unobserved sensitive data X1,,XnX_1,\dots, X_n, we establish local asymptotic mixed normality (along subsequences) of the model Q(n)P=(Q(n)Pθn)θΘQ^{(n)}\mathcal P=(Q^{(n)}P_\theta^n)_{\theta\in\Theta} generating the sanitized observations Z1,,ZnZ_1,\dots, Z_n, where Q(n)Q^{(n)} is an arbitrary sequence of sequentially interactive privacy mechanisms. This result readily implies convolution and local asymptotic minimax theorems. In case p=1p=1, the optimal asymptotic variance is found to be the inverse of the supremal Fisher-Information supQQαIθ(QP)R\sup_{Q\in\mathcal Q_\alpha} I_\theta(Q\mathcal P)\in\mathbb R, where the supremum runs over all α\alpha-differentially private (marginal) Markov kernels. We present an algorithm for finding a (nearly) optimal privacy mechanism Q^\hat{Q} and an estimator θ^n(Z1,,Zn)\hat{\theta}_n(Z_1,\dots, Z_n) based on the corresponding sanitized data that achieves this asymptotically optimal variance.

Keywords

Cite

@article{arxiv.2301.10600,
  title  = {Efficiency in local differential privacy},
  author = {Lukas Steinberger},
  journal= {arXiv preprint arXiv:2301.10600},
  year   = {2024}
}
R2 v1 2026-06-28T08:19:55.063Z