Anonymous Is Harder Than You Think

A clinic deletes every name and publishes its records. You, armed with nothing but a public voter roll, will unmask its patients in three clicks. Then you switch sides and defend the release with k-anonymity, and discover that every unit of privacy is paid for in answers.

1 · The attack: link two harmless tables

You are the attacker. On the left, Midvale's public voter roll: names attached to ZIP, age, and gender, as many real voter files are. On the right, the clinic's "anonymized" release. Pick a neighbor and join the tables on the three shared columns. Whenever exactly one record matches, that person's diagnosis is yours. Re-identify 3 people.

2 · The defense: generalize until k ≥ 5

Now you are the data steward, republishing the same 32 records. Coarsen each column until the smallest group of identical rows holds at least 5 people, and watch what each step does to the attack and to the three queries an analyst still needs to answer. There is more than one way to reach k = 5; they do not cost the same utility.

3 · The tradeoff: privacy is bought with utility

Try at least 3 different settings and compare where they land. Full detail sits at k = 1 with perfect utility; total suppression reaches k = 32 and answers nothing. Real releases live on the frontier between them, and choosing the point on that frontier is an ethical decision, not a technical one: it decides whose privacy is protected and which questions can still be answered.