Workplace privacy lab: k-anonymity and differential privacy

Scenario

Generalise quasi-identifiers

Age band

Department

Location

Data minimisation: drop gender column

Required k

Suppress records in classes smaller than k

Colour people by

Released-table metrics

classes-

k achieved-

unique (raw gen.)-

suppressed-

avg risk-

max risk-

info loss NCP-

utility-

salary exposed-

records-

Risk = 1 / class size (prosecutor model). Utility = (1 - mean certainty penalty) x share of records kept. "Salary exposed" = people in a class whose salaries differ by under $15k: the attacker learns salary without knowing who is who.

Differential privacy: release of average salary

Privacy budget

group-

true mean-

released-

noise scale b-

releases0

total budget used0

Privacy vs utility

Dots: ladder from raw to fully generalised data at the current k and suppression. White marker: your current settings. Lower-right is private but less useful.

Expected DP error vs budget

Laplace mechanism, bounded mean of salary in [$20k, $200k]: sensitivity = range / n, noise scale b = sensitivity / epsilon (Dwork et al. 2006). k-anonymity: Sweeney 2002. All data are synthetic.