Non-homogeneous Anonymizations

Aug 31, 2011

Download: MP4 Video Size: 447.7MB  
Watch on YouTube

Abstract

Privacy Preserving Data Publishing (PPDP) is an evolving research field that is targeted at developing anonymization techniques to enable publishing data so that privacy is preserved while data distortion is minimized. Up until recently most of the research on PPDP considered partition-based anonymization models. The approach in such models is to partition the database records into groups and then homogeneously generalize the quasi-identifiers in all records within a group, as a countermeasure against linking attacks. We describe in this talk alternative anonymization models which are not based on partitioning and homogeneous generalization. Such models extend the set of acceptable anonymizations of a given table, whence they allow achieving similar privacy goals with much less information loss. We shall briefly review the basic models of homogeneous anonymization (e.g. k-anonymity and l-diversity) and then define non-homogeneous anonymization, discuss its privacy, describe algorithms and demonstrate the advantage of such anonymizations in reducing the information loss. We shall then discuss the usefulness of those models for data mining purposes. In particular, we will show that the reduced information loss that characterizes such anonymizations translates also to enhanced accuracy when using the anonymized tables to learn classification models.
 
Based on joint works with Aris Gionis, Arnon Mazza, Mark Last and Sasha Zhmudyak