l-Diversity and t-Closeness
What you'll learn: How l-diversity and t-closeness solve attribute disclosure attacks that k-anonymity cannot prevent.
The Problem with k-Anonymity
You learned that k-anonymity protects against re-identification by ensuring at least k individuals share the same quasi-identifiers. But what if all k people in a group have the same sensitive attribute value?
Imagine a hospital dataset with k=5 anonymity. If all 5 patients in an equivalence class have "Cancer" as their diagnosis, an attacker knowing someone is in that group learns they have cancer—without re-identifying them. This is attribute disclosure, and k-anonymity offers no defense.
l-Diversity: Ensuring Attribute Variety
l-diversity requires each equivalence class to contain at least l "well-represented" distinct values for sensitive attributes.
If l=3, every group of k-anonymous records must have at least 3 different diagnoses (e.g., "Cancer," "Diabetes," "Healthy"). Now an attacker cannot assume everyone in the group shares one condition.
Analogy: k-anonymity creates a crowd where you can't tell who's who. l-diversity ensures the crowd also has diverse characteristics, so knowing someone is in the crowd doesn't reveal their trait.
t-Closeness: Addressing Distribution Skew
Even l-diversity has weaknesses. If 95% of the population has "Healthy" and one equivalence class has 95% "Cancer," the distribution is suspicious and leaks information.
t-closeness requires the distribution of sensitive attributes in each equivalence class to be within distance t of the overall dataset distribution (often using Earth Mover's Distance).
This prevents attackers from learning information through distribution differences, even when values are diverse.
Key Takeaway:
l-diversity ensures sensitive attributes have at least l distinct values per group, preventing homogeneity attacks; t-closeness further requires attribute distributions match the global dataset within threshold t, preventing skewness attacks.