Privacy-Preserving Affinity Propagation Clustering over Vertically Partitioned Data

Abstract

Data mining has been well-studied in academia and widely applied to many fields. As a significant mining means, clustering algorithm has been successfully used in facility location, image categorization and bioinformatics. K-means and affinity propagation (AP) are two effective clustering algorithms, in which the former has involved in privacy preserving data mining, but the latter does not. Considering the unparalleled advantages of AP over k-means, we firstly propose a secure scheme for AP clustering in this paper. Our scheme runs over a partitioned database that different parties contain different attributes for a common set of entities. This scheme guarantees no disclosure of parties' private information by means of the cryptographic tools which have been successfully applied in privacy preserving k-means clustering. The final result for each party is the assignment of each entity, but gives nothing about the attributes held by other parties. In the end, we make a brief security discussion under the semi-honest model and analyze the communication cost to show that our scheme does have good performance.