关于「组合定理与预算会计:基本、高级与最优三种账法」,《差分隐私与经营统计发布》给出的定量结论是什么?
结论:第四章并排计算基本组合、高级组合与最优组合,每天发布一年时最优组合只有基本组合的 0.283。
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。
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: 0.283
| total eps after k releases, delta = 1e-6 | |||||
| eps0 | k | basic | advanced | exact exact/basic | |
| 0.10 | 1 | 0.100 | 0.536 | 0.100 | 1.000 |
| 0.10 | 4 | 0.400 | 1.093 | 0.400 | 1.000 |
| 0.10 | 12 | 1.200 | 1.947 | 1.198 | 0.998 |
| 0.10 | 52 | 5.200 | 4.337 | 3.241 | 0.623 |
| 0.10 | 365 | 36.500 | 13.881 | 10.314 | 0.283 |
| 0.10 | 1000 | 100.000 | 27.140 | 19.345 | 0.193 |
| 0.01 | 1 | 0.010 | 0.053 | 0.010 | 1.000 |
| 0.01 | 4 | 0.040 | 0.106 | 0.040 | 1.000 |
| 0.01 | 12 | 0.120 | 0.183 | 0.116 | 0.968 |
| 0.01 | 52 | 0.520 | 0.384 | 0.274 | 0.526 |
| 0.01 | 365 | 3.650 | 1.041 | 0.790 | 0.216 |
| 0.01 | 1000 | 10.000 | 1.763 | 1.365 | 0.137 |
查看完整验算程序
# Chapter 4. p41: what k releases at eps0 each cost in total.
# Basic composition: k * eps0. Advanced composition (Dwork, Rothblum and
# Vadhan 2010): sqrt(2 k ln(1/d)) * eps0 + k * eps0 * (e**eps0 - 1),
# valid with failure probability d. Exact optimum (Kairouz, Oh and
# Viswanath): randomized response is the worst eps0 mechanism, its loss
# is +eps0 or -eps0 per release, and
# delta(E) = sum_j Binom(k, j) p**j q**(k-j) (1 - e**(E - (2j - k) eps0))+
# with p = e**eps0 / (1 + e**eps0). Bisection finds the least E with
# delta(E) <= d. d = 1e-6 throughout.
import math
D = 1e-6
def advanced(k, e0, d=D):
return math.sqrt(2.0 * k * math.log(1.0 / d)) * e0 + \
k * e0 * (math.exp(e0) - 1.0)
def delta_exact(k, e0, big_e):
lp = e0 - math.log1p(math.exp(e0))
lq = -math.log1p(math.exp(e0))
terms = []
for j in range(k + 1):
loss = (2 * j - k) * e0
if loss <= big_e:
continue
lw = (math.lgamma(k + 1) - math.lgamma(j + 1)
- math.lgamma(k - j + 1) + j * lp + (k - j) * lq)
terms.append(math.exp(lw) * (1.0 - math.exp(big_e - loss)))
return math.fsum(terms)
def exact(k, e0, d=D):
lo, hi = 0.0, k * e0
if delta_exact(k, e0, 0.0) <= d:
return 0.0
for _ in range(80):
mid = 0.5 * (lo + hi)
if delta_exact(k, e0, mid) <= d:
hi = mid
else:
lo = mid
return hi
print("total eps after k releases, delta = 1e-6")
print("%6s %6s %9s %10s %9s %10s"
% ("eps0", "k", "basic", "advanced", "exact", "exact/basic"))
for e0 in (0.1, 0.01):
for k in (1, 4, 12, 52, 365, 1000):
b, a, x = k * e0, advanced(k, e0), exact(k, e0)
print("%6.2f %6d %9.3f %10.3f %9.3f %10.3f"
% (e0, k, b, a, x, x / b))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 4 章「组合定理与预算会计:基本、高级与最优三种账法」(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}
}