Database Reference
In-Depth Information
called DSR, reduces the confidence of the rule by decreasing the support of the
itemset in its right hand side (rule consequent). Both algorithms experience the item
ordering effect under which, based on the order that the sensitive items are hidden,
the produced sanitized databases can be different. Moreover, the DSR algorithm is
usually more effective when the sensitive items have a high support. Compared to
the work of Saygin, et al. [63, 64], the algorithms presented in [76, 77] require a
reduced number of database scans and have an efficient pruning strategy. However,
by construction, they are assigned the task of hiding all the rules containing the sen-
sitive items on their left hand side, while the algorithms in the work of Saygin, et
al. [63, 64] can hide any type of sensitive association rule.
Lee, et al. [42] introduce a data distortion approach that operates by first con-
structing a sanitization matrix from the original data and then multiplying the orig-
inal database (represented as a transactions-by-items matrix) with the sanitization
matrix in order to obtain the sanitized database. The applied matrix multiplication
strategy follows a new definition that aims to enforce the suppression of selected
items from transactions of the original database, thus reduce the support of the
sensitive itemsets. Along these lines, the authors develop three sanitization algo-
rithms: Hidden-First (HF), Non-Hidden-First (NHF) and HPCME (Hiding sensitive
Patterns Completely with Minimum side Effect on nonsensitive patterns). The first
algorithm takes a drastic approach to eliminate the sensitive knowledge from the
original database and is shown to lead to hiding solutions that suffer from the loss
of nonsensitive itemsets. The second algorithm, focuses on the preservation of the
nonsensitive patterns and thus may fail to hide all the sensitive knowledge from the
database. Last, the third algorithm tries to combine the advantages of HF and NHF
in order to hide all sensitive itemsets with minimal impact on the nonsensitive ones.
To achieve this goal, the algorithm introduces a factor restoration probability that it
uses to decide when the preservation of the nonsensitive patterns does not affect the
hiding of the sensitive ones, and thus take the appropriate action.
