Scalable Privacy-Preserving Data Mining with Asynchronously Partitioned Datasets

In the Naive Bayes classification problem using a vertically partitioned dataset, the conventional scheme to preserve privacy of each partition uses a secure scalar product and is based on the assumption that the data is synchronised amongst common unique identities. In this paper, we attempt to discard this assumption in order to develop a more efficient and secure scheme to perform classification with minimal disclosure of private data. Our proposed scheme is based on the work by Vaidya and Clifton[1], which uses commutative encryption to perform secure set intersection so that the parties with access to the individual partitions have no knowledge of the intersection. The evaluations presented in this paper are based on experimental results, which show that our proposed protocol scales well with large sparse datasets.