Frequent itemset mining with differential privacy based on transaction truncation

Abstract

Frequent itemset mining is the basis of discovering transaction relationships and providing information services such as recommendation. However, when transaction databases contain individual sensitive information, direct release of frequent itemsets and their supports might bring privacy risks to users. Differential privacy provides strict protection for users, it can distort the sensitive data when attackers get the sensitive data from statistical information. The transaction length is related to sensitivity for counting occurrences (SCO) in a transaction database, larger SCO will reduce the availability of frequent itemsets under ε-differential privacy. So it is necessary to truncate some long transactions in transaction databases. We propose the algorithm FI-DPTT, a quality function is designed to calculate the optimal transaction length in exponential mechanism (EM), it aims to minimize noisy supports. Experimental results show that the proposed algorithm improves the availability and privacy efficiently.