Generalization
is a well-known method for privacy reserving data publication. Despite its vast
popularity, it has several drawbacks such as heavy information loss, difficulty
of supporting marginal publication, and so on. To overcome these drawbacks, we
develop ANGEL,1 a new anonymization technique that is as effective as
generalization in privacy protection, but is able to retain significantly more
information in the microdata. ANGEL is applicable to any monotonic principles
(e.g., l-diversity, t-closeness, etc.), with its superiority (in correlation
preservation) especially obvious when tight privacy control must be enforced.
We show that ANGEL lends itself elegantly to the hard problem of marginal
publication. In particular, unlike generalization that can release only
restricted marginals, our technique can be easily used to publish any marginals
with strong privacy guarantees.