Distributed Privacy Preserving Classification Based on Local Cluster Identifiers

This paper addresses privacy preserving classification for vertically partitioned datasets. We present an approach based on information hiding that is similar to the basic idea of microaggregation. We use a local clustering to mask the dataset of each party and replace the original attributes by cluster identifiers. That way, the masked datasets can be integrated and used to train a classifier without further privacy restrictions. We apply our approach to four standard machine learning datasets and present the results.