Privacy-Preserving Frequent Itemset Mining for Sparse and Dense Data

Frequent itemset mining is a data mining task that can in turn be used for other purposes such as associative rule mining. The data may be sensitive. There exist multiple privacy-preserving solutions for frequent itemset mining, which should consider the tradeoff between efficiency and privacy. Leaking some less sensitive information such as density of the datatable may improve the efficiency. In this paper, we consider secure multiparty computation setting, where the final output (the frequent itemsets) is public, and no other information should be inferred by the adversary that corrupts some of the computing parties. We devise privacy-preserving algorithms that have advantage when applied to very sparse and very dense matrices. We compare them to related work that has similar security requirements, estimating the efficiency of our new solution on a similar secure multiparty computation platform.