Why does k-anonymity perform poorly on high-dimensional datasets?

With many quasi-identifiers, almost every record becomes unique, so forming equivalence classes of size k requires destroying most of the data.

This is the curse of dimensionality applied to privacy. Each extra quasi-identifier column multiplies the space of possible QID combinations. With enough columns, nearly every individual is unique on their QID vector — so to group k of them together you'd have to generalize/suppress so heavily that little useful information survives.

The result: on wide datasets (think dozens of behavioral attributes), k-anonymity becomes impossible or prohibitively costly to achieve with any meaningful utility. This is partly why modern approaches lean toward differential privacy, federated learning, or synthetic data for high-dimensional data.

Tip: k-anonymity loves short, fat tables (few columns, many rows). It hates long, thin behavioral data where everyone is a snowflake.

Go deeper:

• Curse of dimensionality (Wikipedia) — why records become unique as columns grow.
• Robust De-anonymization (Narayanan & Shmatikov, 2008) — sparsity/high-dimensionality enabling re-identification.