ldiversity

L-diversity, often written as l-diversity, is a privacy criterion used in data anonymization to prevent attribute disclosure. It extends k-anonymity by requiring that, within every group of records that share the same quasi-identifier values (an equivalence class), there are at least l distinct values for the sensitive attribute.

In practice, data publishers generalize or suppress quasi-identifiers to form equivalence classes. A dataset satisfies l-diversity if for every class, the set of values observed for the sensitive attribute contains at least l distinct values. For example, in a hospital dataset with a sensitive attribute "disease," a class with five patients all labeled with different diseases would meet l=3, while a class where all five have the same disease would not.

Benefits of l-diversity include reducing the risk that an attacker who knows the quasi-identifiers can deduce a person’s sensitive attribute from a small set of possible values. It helps mitigate homogeneous attribute disclosure that can occur under k-anonymity alone.

Limitations of l-diversity include its vulnerability to skewness and similarity attacks, where the distribution or semantic relatedness of values can still allow inference even when there are multiple distinct sensitive values in a class. It may also require substantial data generalization or suppression, reducing data utility. More advanced models such as t-closeness and differential privacy address some of these weaknesses. See also k-anonymity, t-closeness, and differential privacy.