Secure distributed framework for achieving ε-differential privacy

Abstract

Privacy-preserving data publishing addresses the problem of disclosing sensitive data when mining for useful information. Among the existing privacy models, ε-differential privacy provides one of the strongest privacy guarantees. In this paper, we address the problem of private data publishing where data is horizontally divided among two parties over the same set of attributes. In particular, we present the first generalization-based algorithm for differentially private data release for horizontally-partitioned data between two parties in the semi-honest adversary model. The generalization algorithm correctly releases differentially-private data and protects the privacy of each party according to the definition of secure multi-party computation. To achieve this, we first present a two-party protocol for the exponential mechanism. This protocol can be used as a subprotocol by any other algorithm that requires exponential mechanism in a distributed setting. Experimental results on real-life data suggest that the proposed algorithm can effectively preserve information for a data mining task. © Springer-Verlag Berlin Heidelberg 2012.