What is a similarity attack, and how can it defeat a dataset that satisfies l-diversity?

A similarity attack exploits semantically related sensitive values: even with l distinct values, if they all mean the same thing, the attacker learns the category.

Similarity attacks target datasets that do satisfy l-diversity by exploiting the semantic proximity of the sensitive values within an equivalence class. Even when values are technically distinct, if they're semantically similar the attacker can narrow the answer to a small, related set.

Worked example — an equivalence class (Age 35–40, ZIP 94102) with four conditions: Gastric Ulcer, Stomach Cancer, Gastritis, Intestinal Disease. This satisfies 4-diversity (four distinct values), yet every value is a digestive-system problem. An attacker confidently concludes: "This individual has a gastrointestinal disorder." — a meaningful, sensitive leak despite "diversity."

Defenses: semantic-distance requirements (use taxonomies/ontologies to force genuinely different values), t-closeness (match the global distribution), and differential privacy (noise-based guarantees independent of semantics).

Tip: Four flavors of the same disease aren't diverse. Real diversity must be measured by meaning, not by counting labels.