What does l-diversity add to k-anonymity, and what is the typical relationship between l and k?

l-diversity requires each equivalence class to hold at least l "well-represented" distinct sensitive values; in practice 1 < l ≤ k.

l-diversity strengthens k-anonymity by attacking the attribute-disclosure blind spot: within every equivalence class, the sensitive attribute must have at least l different, well-represented values, so no single value dominates. This guards against the homogeneity and basic background-knowledge attacks.

Because you can't have more distinct sensitive values than there are records in a class, l is bounded: l is greater than 1 and at most k (1 < l ≤ k). Like k, there's no perfect value of l — it's another risk/utility tuning knob.

Its challenges: it can be mathematically impossible on highly skewed data without huge loss, it reduces utility through aggressive generalization, and finding optimal l-diverse partitions is computationally expensive with many QIDs/sensitive attributes.

Tip: k sets the crowd size; l sets how many different secrets must live in that crowd. You can never need more diversity (l) than the crowd size (k) allows.