Corruption-Resistant Privacy Preserving Distributed EM Algorithm for Model-Based Clustering
Kaleb L. Leemaqz, Sharon Lee, Geoffrey John McLachlan · 2017
Statistical clustering plays an important role in data analysis and is one of the most widely used data mining methods. Concerns about the security and privacy of analyzing modernday massive data across distributed networks have prompted the development of privacy preserving data mining algorithms. This paper proposes a scheme for model-based clustering and classification through a privacy-preserving EM-based learning of a mixture model. We focus on cooperative learning in a multiparty scenario where the parameters of the mixture model can be estimated on the entire data and learnt by all parties without disclosing any private data. In contrast to most existing works which assumed the adversary is Honest-but-Curious, we consider the seldom studied and much stronger and more realistic case of Malicious adversary with unlimited corruption capabilities. The proposed scheme adopts a cyclic communication topology and utilizes cryptographic techniques to encrypt communicated messages, rendering it resistant to multiple corrupted parities. By enforcing one way communication across a ring topology, no trust level hierarchy is required. Upon completion of the training algorithm, each party obtains a clustering of its own private data and is able service a third party by providing predicted cluster labels for new data. For illustration, the Gaussian mixture model is used to present our scheme.