Concept Contrast Fable

Mask sensitive fields and make statistical results hard to reverse-engineer into individuals

Conclusion first:PII masking handles sensitive fields in detailed records, while differential privacy handles the risk of reverse inference in statistical publication or analysis results. One is like covering the names in a ledger, the other like adding controlled perturbation to the general accounts.

PII masking handles sensitive fields in detailed records, while differential privacy handles the risk of reverse inference in statistical publication or analysis results. One is like covering the names in a ledger, the other like adding controlled perturbation to the general accounts.

Fable

Blacked-out names and a perturbed general ledger

The bank wanted to hand a year of accounts to city scholars to study changes in consumption. The bookkeeper first blacked out names, addresses and seals, leaving only age, district, amount and date. The shopkeeper thought this was safe enough, since the most obvious identities were invisible.

As soon as the scholar got the ledger, he pointed out one thing: a seventy-six-year-old coppersmith on East Street had bought three hundred catties of copper the day before the Dragon Boat Festival, and only one person in the city matched. Although the name was crossed out, the person could still be reassembled. Only then did the accountant realize that masking prominent fields only prevents the most straightforward identification.

The second time, the accountant made the blocks coarser, grouped ages into brackets, and merged rare transactions. Individuals were harder to identify, but when scholars asked whether consumption by all elderly people in the city rose before the Dragon Boat Festival, the results could still be swayed by extreme small groups. The cloth had grown thicker, but statistical problems could still leave traces of individuals.

The old statistician offered another approach. Each time, scholars were only allowed to ask aggregate questions, and the bank did not hand over the original ledgers; when answering, it also added a little controlled noise to the totals and limited repeated follow-up questions of the same type. Whether a single person was in the accounts should not let outsiders reliably infer his life trajectory.

The accountant worried the numbers would be inaccurate. The old statistician said: the stronger the protection, the blurrier the details; but if designed well, major trends can still be seen while individuals are not easily reverse-engineered. The first gate is reducing exposed fields; the second gate is controlling how much the answer itself leaks.

The silver house later separated the two doors: before submitting materials, direct identity was masked first to reduce exposure; when publishing statistics, aggregation, perturbation and query rules were used to reduce inference risk. The two doors could be used at the same time, but could not impersonate each other.

Masking names is like draping a cloth over the ledger page; protecting statistics is about controlling the shadow each person leaves when answering questions. One handles the visible marks on the raw page, the other handles others' ability to infer an individual from the answers.

Later, a scholar wanted to follow up repeatedly on consumption changes on the same street. The bank did not open up question-by-question answers, but checked whether these questions together might get too close to a particular person. Only then did the accountant understand that risk is not hidden only in a single ledger page; it can also slowly surface from many seemingly ordinary answers.

Data privacy and PII masking

PII masking reduces the risk of exposing detailed data by deleting, replacing, generalizing, or hiding sensitive fields, and is commonly used in logs, customer service records, training data, and internal analysis.

Differential privacy

Differential privacy uses noise, query restrictions, and privacy budgets to make statistical results difficult to use to reveal whether any individual exists or their specific behavior, making it suitable for statistical release and group analysis.

Story Mapping

Black out names and replace with numbers: PII masking obscures sensitive fields
The timing and amount of rare medicinal materials can still identify customers: re-identification risk remains after masking
Publish only the perturbed totals and averages: differential privacy protects individual information in statistical outputs
Cannot inspect anomalous transactions one by one: differential privacy sacrifices some detail-level usability
Separate procedures for the detail layer and the statistical layer: a real privacy solution requires combining controls by usage scenario

Falls into the supplier privacy commitment

When you hear "de-identification," ask whether it applies to fields, records, or statistical results. PII masking is suitable for reducing detail exposure, but it cannot automatically prevent re-identification; differential privacy is better suited for public statistics and analytics interfaces, but it changes precision and queryability. When procuring AI services, these two types of commitments cannot substitute for each other.

Data privacy and PII masking Differential privacy