Better and Simpler Lower Bounds for Differentially Private Statistical Estimation

被引:0
|
作者
Narayanan, Shyam [1 ]
机构
[1] Citadel Secur, Miami, FL 33131 USA
关键词
Estimation; Complexity theory; Heavily-tailed distribution; Approximation algorithms; Privacy; Differential privacy; Polynomials; Upper bound; Gaussian distribution; Eigenvalues and eigenfunctions; statistics; parameter estimation; lower bounds; fingerprinting;
D O I
10.1109/TIT.2024.3511624
中图分类号
TP [自动化技术、计算机技术];
学科分类号
0812 ;
摘要
We provide optimal lower bounds for two well-known parameter estimation (also known as statistical estimation) tasks in high dimensions with approximate differential privacy. First, we prove that for any alpha <= O(1) , estimating the covariance of a Gaussian up to spectral error ohm (d3/2/alpha epsilon+d/alpha(2)) samples, which is tight up to logarithmic factors. This result improves over previous work which established this for alpha <= O(1/root d) , and is also simpler than previous work. Next, we prove that estimating the mean of a heavy-tailed distribution with bounded kth moments requires ohm (d/alpha(k)/(k-1)epsilon+d/alpha(2)) samples. Previous work for this problem was only able to establish this lower bound against pure differential privacy, or in the special case of k = 2 . Our techniques follow the method of fingerprinting and are generally quite simple. Our lower bound for heavy-tailed estimation is based on a black-box reduction from privately estimating identity-covariance Gaussians. Our lower bound for covariance estimation utilizes a Bayesian approach to show that, under an Inverse Wishart prior distribution for the covariance matrix, no private estimator can be accurate even in expectation, without sufficiently many samples.
引用
收藏
页码:1376 / 1388
页数:13
相关论文
共 50 条
  • [21] Lower Bounds for Quantum Parameter Estimation
    Walter, Michael
    Renes, Joseph M.
    IEEE TRANSACTIONS ON INFORMATION THEORY, 2014, 60 (12) : 8007 - 8023
  • [22] Differentially private resilient distributed cooperative online estimation over digraphs
    Wang, Jimin
    Zhang, Ji-Feng
    Liu, Xiao-Kang
    INTERNATIONAL JOURNAL OF ROBUST AND NONLINEAR CONTROL, 2022, 32 (15) : 8670 - 8688
  • [23] A lower bound on the release of differentially private integer partitions
    Alda, Francesco
    Simon, Hans Ulrich
    INFORMATION PROCESSING LETTERS, 2018, 129 : 1 - 4
  • [24] LOWER AND UPPER BOUNDS ON THE RANDOMNESS COMPLEXITY OF PRIVATE COMPUTATIONS OF AND
    Kushilevitz, Eyal
    Ostrovsky, Rafail
    Prouff, Emmanuel
    Rosen, Adi
    Thillard, Adrian
    Vergnaud, Damien
    SIAM JOURNAL ON DISCRETE MATHEMATICS, 2021, 35 (01) : 465 - 484
  • [25] Asymptotically Optimal and Private Statistical Estimation
    Smith, Adam
    CRYPTOLOGY AND NETWORK SECURITY, PROCEEDINGS, 2009, 5888 : 53 - 57
  • [26] SAMPLE COMPLEXITY BOUNDS ON DIFFERENTIALLY PRIVATE LEARNING VIA COMMUNICATION COMPLEXITY
    Feldman, Vitaly
    Xiao, David
    SIAM JOURNAL ON COMPUTING, 2015, 44 (06) : 1740 - 1764
  • [27] Towards sharper excess risk bounds for differentially private pairwise learning
    Kang, Yilin
    Li, Jian
    Liu, Yong
    Wang, Weiping
    NEUROCOMPUTING, 2024, 610
  • [28] Statistic selection and MCMC for differentially private Bayesian estimation
    Barış Alparslan
    Sinan Yıldırım
    Statistics and Computing, 2022, 32
  • [29] Differentially private estimation in a class of bipartite graph models
    Pan, Lu
    Hu, Jianwei
    COMMUNICATIONS IN STATISTICS-THEORY AND METHODS, 2024, 53 (18) : 6477 - 6496
  • [30] Gaussian differentially private robust mean estimation and inference
    Yu, Myeonghun
    Ren, Zhao
    Zhou, Wen-xin
    BERNOULLI, 2024, 30 (04) : 3059 - 3088