关于「敏感度:营收先截断,客群分布只花一次预算」,《差分隐私与经营统计发布》给出的定量结论是什么?

结论: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。

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: 4.09, 6545

true monthly revenue 159869 yuan from 400 customers, eps = 0.5
Cclippedbiasnoise sdRMSERMSE %
2500.560789717077897449.40
5000.2424177514144179926.15
7500.1102565221212573916.10
10000.0681684528281708010.68
15000.0307505424386215.39
20000.0133292565765454.09
30000.005970848585415.34
50000.000014142141428.85
100000.0000282842828417.69
best of the grid: C = 2000, RMSE 6545 yuan
spend quantiles: p50 278p90 789p99 2086max 3623
if one customer may add up to 6 orders, order-level clipping
at C = 2000 needs noise sd 33941, not 5657
查看完整验算程序
# Chapter 3. p31: choosing the clipping bound C for a revenue total.
# One customer can change an unclipped revenue sum by any amount, so its
# sensitivity is unbounded. Clip each customer's monthly spend at C: the
# sensitivity becomes C, the noise sd is sqrt(2) * C / eps, and the price
# is a bias, the spend above C that is cut away. Root mean square error
# = sqrt(bias**2 + 2 * C**2 / eps**2). 400 customers with log-normal
# spend (median 300 yuan, sigma 0.9; illustrative), eps = 0.5.
import math
import random

rng = random.Random(20261113)
spend = [math.exp(rng.gauss(math.log(300.0), 0.9)) for _ in range(400)]
total = math.fsum(spend)
EPS = 0.5
print("true monthly revenue %.0f yuan from %d customers, eps = %.1f"
      % (total, len(spend), EPS))
print("%6s %8s %10s %10s %10s %8s"
      % ("C", "clipped", "bias", "noise sd", "RMSE", "RMSE %"))
best = None
for c in (250, 500, 750, 1000, 1500, 2000, 3000, 5000, 10000):
    cut = math.fsum(max(0.0, x - c) for x in spend)
    share = len([x for x in spend if x > c]) / len(spend)
    sd = math.sqrt(2.0) * c / EPS
    rmse = math.sqrt(cut * cut + sd * sd)
    if best is None or rmse < best[1]:
        best = (c, rmse)
    print("%6d %8.3f %10.0f %10.0f %10.0f %8.2f"
          % (c, share, cut, sd, rmse, 100.0 * rmse / total))
print("best of the grid: C = %d, RMSE %.0f yuan" % best)
q = sorted(spend)
print("spend quantiles: p50 %.0f  p90 %.0f  p99 %.0f  max %.0f"
      % (q[199], q[359], q[395], q[-1]))
# a customer with k orders counted as k rows has sensitivity k * C
print("if one customer may add up to 6 orders, order-level clipping")
print("at C = %d needs noise sd %.0f, not %.0f"
      % (best[0], 6 * math.sqrt(2.0) * best[0] / EPS,
         math.sqrt(2.0) * best[0] / EPS))

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 3 章「敏感度:营收先截断,客群分布只花一次预算」(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