关于「拉普拉斯机制:ε 是似然比的天花板」,《差分隐私与经营统计发布》给出的定量结论是什么?

结论:第二章给出拉普拉斯机制与 ε 的定义:计数查询取 ε 为 1 时 95 % 误差半宽约 2.996 人,任何对手比瞎猜最多好 0.4621。

数据口径:书中参数为示意;全部计算由书中程序生成;本书是非个人化的法律与经营建议,差分隐私是数学保证,不替代法律合规,法规只述大意,以上架前人工复核为准。 第三章处理敏感度:营收按人截断,示意数据中均方根误差 6545 元,占营收 4.09 %;客群分布按并行组合只花一次预算。

English version

对照数据表

下表数字由本章 Python 程序原样打印,可重新运行复现。 程序只用 Python 标准库与结论吻合的打印值: 2.996, 95

count query, sensitivity 1, b = 1 / eps
epsbmean|X|2b^2var95% hwP(>hw)
0.110.00010.023200.000201.24629.9570.0503
0.52.0002.0008.0008.0055.9910.0506
1.01.0000.9982.0001.9892.9960.0493
2.00.5000.5000.5000.5011.4980.0504
99% half-width at eps 1: 4.605; 99.9%: 6.908
查看完整验算程序
# Chapter 2. p22: how large Laplace noise is, in closed form and by draws.
# Laplace(b) with b = sensitivity / eps: mean |X| = b, variance 2 b**2,
# P(|X| > t) = exp(-t / b). So with probability 1 - beta the error is at
# most b * ln(1 / beta). Sampling by inverse CDF from a seeded generator;
# 200000 draws per row. Sensitivity 1 (a count of people).
import math
import random

rng = random.Random(20261113)
DRAWS = 200000


def laplace(r, b):
    u = r.random()
    while u == 0.0:
        u = r.random()
    if u < 0.5:
        return b * math.log(2.0 * u)
    return -b * math.log(2.0 * (1.0 - u))


print("count query, sensitivity 1, b = 1 / eps")
print("%5s %9s %9s %9s %9s %9s %9s"
      % ("eps", "b", "mean|X|", "2b^2", "var", "95% hw", "P(>hw)"))
for eps in (0.1, 0.5, 1.0, 2.0):
    b = 1.0 / eps
    xs = [laplace(rng, b) for _ in range(DRAWS)]
    mabs = math.fsum(abs(x) for x in xs) / DRAWS
    mean = math.fsum(xs) / DRAWS
    var = math.fsum((x - mean) ** 2 for x in xs) / DRAWS
    hw = b * math.log(20.0)
    over = len([x for x in xs if abs(x) > hw]) / DRAWS
    print("%5.1f %9.3f %9.3f %9.3f %9.3f %9.3f %9.4f"
          % (eps, b, mabs, 2 * b * b, var, hw, over))
print("99%% half-width at eps 1: %.3f; 99.9%%: %.3f"
      % (math.log(100.0), math.log(1000.0)))

唯一数据源与引用

范雅福(Fan, Y.). (2026). 《差分隐私拉普拉斯机制与隐私预算会计、公开经营统计而不泄露个人的发布母典》第 2 章「拉普拉斯机制:ε 是似然比的天花板」(版本 1.0.0). 5D Library. DOI 待发布

DOI 待发布 · 专著页正式发布后上线(本站暂无专著页) · 5DL-B-296 · v1.0.0 · CC-BY-NC-ND-4.0

BibTeX
@book{fan2026296,
  author    = {Fan, Yafu},
  title     = {Differential Privacy, the Laplace Mechanism and Privacy-Budget Accounting: Publishing Business Statistics without Exposing Any Individual},
  year      = {2026},
  publisher = {5D Library},
  edition   = {1.0.0},
  note      = {DOI pending}
}

本书其他章节结论