Publish useful data.
Limit what it reveals.
Experiment with quasi-identifiers, equivalence classes, disclosure attacks, k-anonymity, l-diversity, and t-closeness.
Three attribute roles
Removing names does not guarantee anonymity. Attackers can link combinations such as ZIP code, birth date, and sex to external records.
Key attribute
Uniquely identifies a person directly.
Name · address · phone numberQuasi-identifier
May identify someone when combined with external information.
ZIP · age · sexSensitive attribute
Needed for analysis but capable of causing privacy harm.
Diagnosis · wage · medicationClassify each attribute
For every row, choose one role. You can change an answer before checking.
The k-anonymity workbench
Every released record must share its quasi-identifier pattern with at least k−1 other records. Increase k to see broader generalization and lower re-identification probability.
Generalization
Replace a value with a less specific, semantically consistent range or category.
Age 27 → 25–29Suppression
Remove selected values and replace them with a wildcard.
ZIP 53711 → 5371*Trade-off
Higher k usually improves protection from record linkage while reducing analytical precision.
Greedy partitioning
A k-d-tree style algorithm repeatedly splits a multidimensional data set while ensuring that each resulting partition can still contain at least k records.
| Step | Partition | Split | Reason |
|---|---|---|---|
| 1 | Full table | ZIP at 53711 | Creates two valid groups |
| 2 | Left group | Age at 26 | Both children retain at least k=2 |
| 3 | Each child | No allowable cut | Another split would create a group smaller than k |
What k-anonymity does not solve
k-anonymity limits identification through quasi-identifiers. It does not guarantee that sensitive values inside an equivalence class remain uncertain.
Everyone has the same diagnosis
A 5-anonymous class still reveals a diagnosis when all five sensitive values match.
Outside facts narrow the answer
Knowing that Carl does not have heart disease may reveal which remaining diagnosis applies.
External records restore identity
Pseudonyms alone can fail when dates, ratings, locations, or demographics overlap with public data.
Attack detector
A 3-anonymous class contains three people. Every sensitive value is “Diabetes.” What remains exposed?
Beyond k-anonymity
Select a model to compare the protection goal and its main limitation.
Distribution thought experiment
HIV prevalence in the full table is 1%. An equivalence class contains 50% positive records and 50% negative records. It has two distinct sensitive values.
Why is diversity alone misleading? The class distribution is far from the population distribution, so membership in the class sharply increases the inferred probability of a positive result. t-closeness evaluates this shift.