Differentially Private Frequent Itemset Mining Against Incremental Updates

Abstract

Differential privacy has recently been applied to frequent itemset mining (FIM). Most existing works focus on promoting result utility while satisfying differential privacy. However, they all focus on “one-shot” release of a static dataset, which do not adequately address the increasing need for up-to-date sensitive information. In this paper, we address the problem of differentially private FIM for dynamic datasets, and propose a scheme against infinite incremental updates which satisfies $\epsilon$-differential privacy in any sliding window. To reduce the increasing perturbation error against incremental updates, we design an adaptive budget allocation scheme combining with transactional dataset change. To reduce the high sensitivity of one-shot release, we split long transactions and analyze its information loss. Then we privately compute the approximate number of frequent itemsets. Based on the above results, we design a threshold exponential mechanism to privately release frequent itemsets. Through formal privacy analysis, we show that our scheme satisfies $\epsilon$-differential privacy in any sliding window. Extensive experiment results on real-world datasets illustrate that our scheme achieves high utility and efficiency.