Inferential or Differential: Privacy Laws Dictate

TL;DR

This paper critically analyzes differential privacy, showing it doesn't ensure inferential privacy and questions the practical impact of known impossibility results, while proposing a flexible, utility-preserving solution for inferential privacy.

Contribution

It provides a critical analysis of differential privacy's limitations and introduces a new approach to inferential privacy that overcomes previous challenges.

Findings

Differential privacy does not guarantee inferential privacy.

Impossibility results depend on adversary models, with uncertain practical implications.

Proposed solution improves flexibility, utility, and robustness against auxiliary information.

Abstract

So far, privacy models follow two paradigms. The first paradigm, termed inferential privacy in this paper, focuses on the risk due to statistical inference of sensitive information about a target record from other records in the database. The second paradigm, known as differential privacy, focuses on the risk to an individual when included in, versus when not included in, the database. The contribution of this paper consists of two parts. The first part presents a critical analysis on differential privacy with two results: (i) the differential privacy mechanism does not provide inferential privacy, (ii) the impossibility result about achieving Dalenius's privacy goal [5] is based on an adversary simulated by a Turing machine, but a human adversary may behave differently; consequently, the practical implication of the impossibility result remains unclear. The second part of this work is devoted to a solution addressing three major drawbacks in previous approaches to inferential privacy: lack of flexibility for handling variable sensitivity, poor utility, and vulnerability to auxiliary information.