跳到正文
DEV Community· Lv Shao Mu·· 15 天前AI 評分53

piimask 如何遮罩 LLM 資料集並保留校驗有效性與關聯

I masked a dataset for an LLM and kept every checksum valid — here's the approach

AI 導讀

Lv Shao Mu 介紹 piimask 的 LLM 資料遮罩方法,以確定性假值取代原始個人資料,同時保留欄位格式、校驗有效性及跨資料列關聯。映射以 HMAC-SHA256 和密鑰生成,同一密鑰及輸入會得到相同假值;工具會先驗證中國居民身份證及銀行卡的校驗,再遮罩通過驗證的值。

正文

Sharing data with an LLM is a data-leak problem with an extra twist: the model is a reader, not a database. If you scrub a spreadsheet with ***, the model has nothing to reason about. If you replace every cell with random noise, joins break and the analysis collapses.

This post walks through the middle path: generate fake values that keep the original shape, so the file still imports, validates, and joins — without any real value surviving.

What breaks with regex-replace

Take a customer export with a phone, a Chinese resident ID, and a bank card:

name,phone,id_no,bank,email
张三,13812345678,11010519491231002X,6222025644974749,zhangsan@example.com

A naive scrubber produces:

name,phone,id_no,bank,email
***,***********,***,***,***

Four failures at once:

  1. Importers reject the row. Phone columns have length checks.
  2. Checksums die. A Chinese ID's last digit is an ISO 7064 MOD 11-2 check character; a bank card ends with a Luhn digit. Random digits fail both.
  3. Joins break. If two rows referenced the same customer by phone, they now reference two different asterisk blobs.
  4. Reruns disagree. Fuzz each run separately and two environments get two different datasets.

The fix: fake values that pass validation

Mask by generating a deterministic fake per kind:

name,phone,id_no,bank,email
张三,13893525401,494626517976288038,6222025644974749,u72765aad12@example.com

Wait — the bank card looks identical. That's the point. The fake preserves:

  • Phone: 11 digits, 1[3-9] prefix.
  • CN ID: 17 digits plus a valid ISO 7064 check character. 198436564585937260 passes sum(digit[i] * W[i]) % 11 against the same table the importer uses.
  • Bank card: same length, same BIN prefix (622202), and a Luhn-valid check digit — the payment sandbox accepts it.
  • Email: hashed local part, same domain, so you can still count distinct senders per domain.
  • IPv4: mapped into 10.x.x.x private space.

And the whole mapping is keyed:

fake = HMAC-SHA256(secret, "kind:value") → truncated into the value's own alphabet

Same secret + same input → same output. Different secret → different output. Without the secret, the published file alone doesn't let you recover the mapping.

Why checksum validation matters more than it sounds

The most common complaint about PII scrubbers is false positives. \d{18} matches order numbers. \d{16} matches SKUs. The difference between "88 real findings" and "400 red herrings" is whether you validate before flagging:

  • a real CN ID satisfies its check character;
  • a real bank card satisfies Luhn;
  • an 18-digit order number satisfies neither.

Validate first, mask only what passes. Skip validation only with an explicit --loose flag, which exists to catch forged numbers — and says so in its output.

CLI walkthrough

Single file, zero dependencies beyond the Python standard library:

git clone https://github.com/lvshaomu/piimask
cd piimask

# one file, with a report
python piimask.py --in customers.csv --out clean/ --report

# a folder, recursively
python piimask.py --in exports/ --out clean/

# keep the shape instead of generating fakes
python piimask.py --in dump.json --out clean/ --policy mask

Output:

     8  sample.csv -> clean/sample.masked.csv

piimask report
========================================
policy        : fake
detectors     : email, id_card_cn, phone_cn, bank_card, ipv4
replacements  : 8
by kind       : email=2, id_card_cn=2, ipv4=2, phone_cn=2

Exit codes matter for CI: 0 ran fine, 2 means --strict found a file with zero replacements. A silent no-op is a failure, not a success.

Tests

The repo ships 18 assertions, including the ones that matter:

PASS  构造的身份证通过 ISO 7064
PASS  掩码后身份证仍通过校验: 198436564585937260
PASS  掩码后银行卡通过 Luhn: 6222025644974749
PASS  掩码后银行卡保留 BIN 前缀
PASS  同 secret 重跑结果一致
PASS  不同 secret 输出不同
PASS  同值两次掩码一致(可 JOIN)
PASS  CLI 掩码后原身份证不再出现
PASS  JSON 结构保留(键仍在)

The two-sided property — "originals gone, fakes still valid" — is what makes a masked file usable downstream.

Where this fits in an LLM workflow

A practical pattern for sharing data with a model:

  1. Mask first with a deterministic keyed mapping (--secret-file key).
  2. Send the masked file — the model sees realistic structure.
  3. Reason about joins — the same customer maps to the same fake everywhere, so "count orders per customer" still works.
  4. Restore only if you must — keep the secret; without it, the output is not reversible.

The failure mode this avoids: scrubbing to *** and then wondering why the model can't tell which rows belong to the same account.

Limits

  • The detectors are shape + checksum based. Free-text Chinese names are deliberately not guessed, because that's exactly how false positives happen. Column headers drive name detection in the larger variant.
  • .xlsx and .docx need a slightly bigger tool; this repo stays text/CSV/JSON on purpose, standard library only.
  • Deterministic ≠ anonymous against someone who has both the file and the secret. Keep the secret with the restore duty.

The repo: lvshaomu/piimask — MIT, single file, no dependencies. A fuller variant with .xlsx/.docx and four policies lives at csvtriage.com/p/datamask.html.

來源:DEV Community · dev.to