Proj. Unknown: Deciding Differential Privacy of Online Algorithms with Multiple Variables

发布时间 2023-11-03 18:43:27作者: 雪溯

Paper

https://arxiv.org/abs/2309.06615

Abstract

背景:

  1. 自动机A被称作查分隐私自动机:当对某些D,对任何隐私预算ε>0,该自动机是Dε-differentially private( A DiP automaton is a parametric automaton whose behavior depends on the privacy budget ?. An automaton A will be said to be differentially private if, for some ?, the automaton is ??-differentially private for all values of ? > 0)

本文:Dip automata(DipA)
对象:处理输入流,并且会为每个输入都产生某种输出的在线算法
方法:

  1. 定义名为Dip Automata的自动机来描述这类允许多个实值存储量的算法(allowing multiple real-valued storage variable)
  • Dip automaton是一类行为依据于隐私预算privacy budget ε的参数自动机
  1. 基于精度特征定义一类差分隐私Dip自动机(Q:还是所有差分隐私Dip自动机,We show that the problem of determining if a given DiP automaton belongs to this class is PSPACE-complete.),证明判定给定的Dip自动机是否属于这类差分隐私Dip自动机的问题为PSPACE-Complete。
  2. 本文中提到的PSPACE-Complete算法也能够计算差分隐私Dip自动机的D值

实验:证明算法确实有效