Paper summary : K-Anonymous Data Mining, A Survey

Extracting useful information from large datasets has become an increasingly common endeavour. However, if datasets are not privacy protected, data mining can lead to a breach of individual privacy. One privacy preserving technique, k-anonymity, is examined. The authors describe k-anonymity, its enforcement, threats to k-anonymity and addressing these threats (in the context of data mining).

k-anonymity requires that the size of a QI-group (equivalence class) is at least k. QI attributes are those that can serve as an identifier for an individual. The aim is for tuples in a QI-group to be indistinguishable from each other. QI-groups can be built by generalization (replacing specific values with generalized versions from a hierarchy) and suppression (changing a specific attribute value to its most generalized form). These techniques result in information loss, which algorithms seek to minimize e.g., Samarati’s algorithms, etc.

k-anonymity can be violated by data mining techniques (e.g., association rule mining, classification, etc) since they can used to draw inferences and breach privacy. To address this, one could either anonymize-and-mine (AM) or mine-and-anonymize (MA). The AM approach decouples anonymization and mining but reduces the utility of the data while the MA approach extracts more utility from the data and is more efficient but restricts data mining to the owner.

To anonymize data in the AM approach, one could either apply top-down or bottom-up algorithms. Top-down algorithms initially generalize tables completely and subsequently refine them, increasing data utility but decreasing anonymity (vice versa for bottom-up).

To anonymize mined output in the MA approach, the algorithms depend on the specific scenario, but the main theme is to detect and close inference channels.

Paper : K-Anonymous Data Mining, A Survey; Ciriani et. Al., Privacy-Preserving Data Mining, Springer-Verlag, 2007