224 ◾ Secure Data Provenance and Inference Control
using decision trees. e assumption is that when Low needs information for pur-
poses such as performance and functionality, High must decide whether to give
(i.e., downgrade) information to Low. In other words, when High wishes to down-
grade a set of data to Low, it may be necessary, because of inference channels, to
trim the set. Basically decision trees are used to form rules from the downgraded
data High makes available to Low. Remember that we can use the nonsensitive
attributes of an individual to arrive at (i.e., predict) the sensitive attribute using
rules that are trained on similar individuals (occurring in previous released data).
In parsimonious downgrading, a cost measure is a