关于「一个指纹的盲区:总哈希只能回答「变没变」」,《默克尔树与书稿存证》给出的定量结论是什么?

结论:第一章在九个文件的示意书稿上做单字节改动,凭总哈希猜改了哪个文件,命中率 0.1133,与瞎猜的 0.1111 无异。

数据口径:书中的书稿、时间戳、盐值与参数均为示意,全部计算由书中程序生成;哈希与时间戳只证明某段内容不晚于某时已经存在且未被改动,不等于著作权登记或权属证明,举证效力以法院与当地法规为准;本书不构成个人化法律建议。 第一章在九个文件的示意书稿上做单字节改动,凭总哈希猜改了哪个文件,命中率 0.1133,与瞎猜的 0.1111 无异。

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: 0.1111, 0.1133

file mean bits sd bits (of 256 output bits, 300 edits each)
0128.09 7.93
1128.15 8.42
2127.96 8.53
3127.95 7.93
4127.83 7.58
5128.11 7.83
6127.70 8.03
7127.69 8.16
8127.47 8.17
grand mean distance127.88 bits
spread of file means0.68 bits
guess-the-file accuracy0.1133 (test edits 2700)
chance level 1/90.1111
standard error at chance0.0060
bits needed to name the file3.1699
查看完整验算程序
# Chapter 1. p11: one digest of the whole book says "changed", not where.
# Model: a demo manuscript of 9 files (preface + 8 chapters) is hashed
# as one SHA-256 digest. We change one byte in a random file at a random
# offset and measure the Hamming distance between old and new digests.
# If the digest carried any trace of WHERE the edit happened, the mean
# distance would differ by file and a classifier could guess the file.
# Training and test edits use different fixed seeds. Texts are synthetic
# (illustrative texts, not a real manuscript).
import hashlib
import math
import random

NFILE = 9


def file_text(i, ver=1):
    # illustrative manuscript: about 5 KB of fixed text per file
    line = "B-295 demo manuscript, file %02d, version %d. " % (i, ver)
    return (line * 120).encode("ascii")


def book_digest(files):
    h = hashlib.sha256()
    for f in files:
        h.update(f)
    return h.digest()


def hamming(a, b):
    return bin(int.from_bytes(a, "big") ^ int.from_bytes(b, "big")).count(
        "1")


def edits(offset, per_file):
    rng = random.Random(20261113 + offset)
    base = [file_text(i) for i in range(NFILE)]
    d0 = book_digest(base)
    out = []
    for i in range(NFILE):
        for _ in range(per_file):
            f = bytearray(base[i])
            pos = rng.randrange(len(f))
            f[pos] = (f[pos] + 1 + rng.randrange(255)) % 256
            files = base[:i] + [bytes(f)] + base[i + 1:]
            out.append((i, hamming(d0, book_digest(files))))
    return out


train = edits(0, 300)
test = edits(1, 300)
means = []
print("file  mean bits  sd bits  (of 256 output bits, 300 edits each)")
for i in range(NFILE):
    xs = [d for f, d in train if f == i]
    m = math.fsum(xs) / len(xs)
    sd = math.sqrt(math.fsum((x - m) ** 2 for x in xs) / (len(xs) - 1))
    means.append(m)
    print("%4d  %9.2f  %7.2f" % (i, m, sd))
hits = 0
for f, d in test:
    guess = min(range(NFILE), key=lambda j: (abs(d - means[j]), j))
    if guess == f:
        hits += 1
grand = math.fsum(d for _, d in train) / len(train)
print("grand mean distance           %.2f bits" % grand)
print("spread of file means          %.2f bits" % (max(means) - min(means)))
print("guess-the-file accuracy       %.4f (test edits %d)"
      % (hits / len(test), len(test)))
print("chance level 1/9              %.4f" % (1 / NFILE))
print("standard error at chance      %.4f"
      % math.sqrt((1 / NFILE) * (1 - 1 / NFILE) / len(test)))
print("bits needed to name the file  %.4f" % math.log2(NFILE))

唯一数据源与引用

范雅福(Fan, Y.). (2026). 《默克尔树哈希链与可信时间戳、专著版本防篡改与优先权举证最小证明母典》第 1 章「一个指纹的盲区:总哈希只能回答「变没变」」(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-295 · v1.0.0 · CC-BY-NC-ND-4.0

BibTeX
@book{fan2026295,
  author    = {Fan, Yafu},
  title     = {Merkle Hash Trees and Trusted Timestamps: Tamper-Evidence for Monograph Versions and Minimal Proofs of Priority},
  year      = {2026},
  publisher = {5D Library},
  edition   = {1.0.0},
  note      = {DOI pending}
}

本书其他章节结论