Database Reference
In-Depth Information
Chapter 4
Other Knowledge Hiding Methodologies
Association rule hiding algorithms aim at protecting sensitive knowledge captured
in the form of frequent itemsets or association rules. However, (sensitive) knowledge
may appear in various forms directly related to the applied data mining algorithm
that achieved to expose it. Consequently, a set of hiding approaches have been pro-
posed over the years to allow for the safeguarding of sensitive knowledge exposed
by data mining tasks such as clustering, classification and sequence mining. In this
chapter, we briefly discuss some state-of-the-art approaches for the hiding of sensi-
tive knowledge that is depicted in any of the aforementioned formats.
4.1 Classification Rule Hiding
Classification rule hiding has been studied to a substantially lesser extent than asso-
ciation rule hiding. Similarly to association rule hiding methodologies, classification
rule hiding algorithms consider a set of classification rules as sensitive and aim to
protect them. Research in the area of classification rule hiding has developed along
two main directions: suppression-based techniques and reconstruction-based tech-
niques. Suppression-based techniques aim at reducing the confidence of a sensitive
classification rule (measured in terms of the owner's belief regarding the holding
of the rule when given the data) by distorting the values of certain attributes in the
database that belong to transactions related to the existence of the rule.
Chang & Moskowitz [15] were the first to address the inference problem caused
by the downgrading of the data in the context of classification decision rules.
Through a blocking technique, called parsimonious downgrading, the authors block
the inference channels that lead to the identification of the sensitive classification
rules by selectively modifying transactions so that missing values appear in the re-
leased database. This has as an immediate consequence the lowering of the confi-
dence regarding the holding of the sensitive rules.
Wang, et al. [74] propose a heuristic approach that achieves to fully eliminate all
the sensitive inferences, while effectively handling overlapping rules. The algorithm
