Tight Bounds for Gaussian Mean Estimation under Personalized Differential Privacy

TL;DR

This paper addresses Gaussian mean estimation under personalized differential privacy (PDP), covering both bounded and unbounded neighbor relations. It develops tight minimax lower bounds and near-matching upper bounds, using a novel diffusion primitive to achieve privacy amplification with nonuniform PDP rates in the bounded setting, and a privacy-vector sanitization plus dataset shrinking approach to handle unbounded PDP. The core contributions are a diffusion-based framework for bounded PDP range estimation and mean estimation, a privacy-specific Le Cam-style lower bound for unbounded PDP, and a reduction that transfers unbounded PDP problems to bounded PDP instances without compromising privacy. The results yield instance-optimal, polylogarithmically tight guarantees for Gaussian mean estimation under PDP, with practical implications for private data analysis when individual privacy preferences vary, including protection of participation information in the unbounded model.

Abstract

We study mean estimation for Gaussian distributions under \textit{personalized differential privacy} (PDP), where each record has its own privacy budget. PDP is commonly considered in two variants: \textit{bounded} and \textit{unbounded} PDP. In bounded PDP, the privacy budgets are public and neighboring datasets differ by replacing one record. In unbounded PDP, neighboring datasets differ by adding or removing a record; consequently, an algorithm must additionally protect participation information, making both the dataset size and the privacy profile sensitive. Existing works have only studied mean estimation over bounded distributions under bounded PDP. Different from mean estimation for distributions with bounded range, where each element can be treated equally and we only need to consider the privacy diversity of elements, the challenge for Gaussian is that, elements can have very different contributions due to the unbounded support. we need to jointly consider the privacy information and the data values. Such a problem becomes even more challenging under unbounded PDP, where the privacy information is protected and the way to compute the weights becomes unclear. In this paper, we address these challenges by proposing optimal Gaussian mean estimators under both bounded and unbounded PDP, where in each setting we first derive lower bounds for both problems, following PDP mean estimators with the algorithmic upper bounds matching the corresponding lower bounds up to logarithmic factors.

Theorems & Definitions (48)

• Theorem 1: Informal, see Theorem \ref{['theorem:lower_bound']}
• Theorem 2: Informal, see Theorem \ref{['theorem:upper_bound_bounded']}
• Theorem 3: Informal, see Theorem \ref{['theorem:lower_bound_add/remove-one']}
• Theorem 4: Informal, see Theorem \ref{['theorem:upper_bound_add/remove-one']}
• Definition 1: Differential Privacy dwork2014algorithmic
• Lemma 1: Post-Processing
• Lemma 2: Basic Composition
• Lemma 3: Parallel Composition Parallel_Composition
• Lemma 4: Laplace Mechanism
• Lemma 5: Laplace Tail Bound