关于「东西方原典考辨:韩非《说难》与沃纳的随机化回答」,《差分隐私与经营统计发布》给出的定量结论是什么?
结论:第七章考辨《韩非子·说难》「语及所匿之事」与沃纳 1965 年的随机化回答(转盘概率 0.8 对应 ε 为 1.386)。
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: 0.8, 1.386
| root mean square error, n | 500, true share 0.3 |
| design p 0.6 p 0.7 p 0.8 p 0.9 | |
| Warner eps = ln(p/q) | 0.405 0.847 1.386 2.197 |
| Warner RMSE | 0.1114 0.0552 0.0362 0.0265 |
| direct, T | 1.00 0.0205 |
| direct, T | 0.95 0.0252 |
| direct, T | 0.90 0.0360 |
| direct, T | 0.80 0.0630 |
| direct, T | 0.60 0.1212 |
| simulated Warner at p 0.8: mean 0.3012, sd 0.0360 |
查看完整验算程序
# Chapter 7. p72: Warner's 1965 trade - variance for honesty.
# Warner's device: with probability p the respondent answers "I belong
# to group A", otherwise "I do not belong to group A", and only she
# knows which. Estimate: pi_hat = (lambda_hat - (1 - p)) / (2 p - 1),
# variance pi (1 - pi) / n + p (1 - p) / (n (2 p - 1)**2). Direct
# questioning is cheaper in variance but biased when members of A tell
# the truth only with probability T: bias = -pi (1 - T). Root mean square
# errors for n = 500 and pi = 0.3 (illustrative); eps = ln(p / (1 - p))
# is the local privacy level that Warner's p implies.
import math
import random
N, PI = 500, 0.3
PS = [0.6, 0.7, 0.8, 0.9]
def rmse_rr(p):
var = PI * (1 - PI) / N + p * (1 - p) / (N * (2 * p - 1) ** 2)
return math.sqrt(var)
def rmse_direct(t):
seen = PI * t
return math.sqrt((PI - seen) ** 2 + seen * (1 - seen) / N)
print("root mean square error, n = %d, true share %.1f" % (N, PI))
print("%-22s" % "design" + "".join("%9s" % ("p %.1f" % p) for p in PS))
print("%-22s" % "Warner eps = ln(p/q)" + "".join(
"%9.3f" % math.log(p / (1 - p)) for p in PS))
print("%-22s" % "Warner RMSE" + "".join("%9.4f" % rmse_rr(p) for p in PS))
for t in (1.0, 0.95, 0.9, 0.8, 0.6):
print("%-22s" % ("direct, T = %.2f" % t) + "%9.4f" % rmse_direct(t))
rng = random.Random(20261113)
p, ests = 0.8, []
for _ in range(4000):
yes = 0
for _ in range(N):
a = rng.random() < PI
say = a if rng.random() < p else not a
if say:
yes += 1
ests.append((yes / N - (1 - p)) / (2 * p - 1))
m = math.fsum(ests) / len(ests)
sd = math.sqrt(math.fsum((e - m) ** 2 for e in ests) / (len(ests) - 1))
print("simulated Warner at p 0.8: mean %.4f, sd %.4f" % (m, sd))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 7 章「东西方原典考辨:韩非《说难》与沃纳的随机化回答」(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}
}