Mining Frequent Itemsets in Presence of Malicious Participants

Privacy Preserving Data Mining (PPDM) algorithms attempt to reduce the injuries to privacy caused by malicious parties during the rule mining process. Usually, these algorithms are designed for the semi-honest model, where participants do not deviate from the protocol. However, in the real-world, malicious parties may attempt to obtain the secret values of other parties by probing attacks or collusion. In this paper we study how we can preserve the privacy of participants in a collusion-free model of the frequent itemset mining process, where the protocol protects against probing attacks and collusion. The mining of frequent itemsets is the main step of association rule mining algorithms, and, in this paper, we propose two privacy-preserving frequent itemset mining algorithms for both twoparty and multi-party states in a collusion-free model for vertically partitioned (heterogeneous) data; in addition, we propose a new privacy measuring technique, which quantifies privacy based on the amount of disclosed sensitive information.