Differentially private approximate aggregation based on feature selection

• Published in: Journal of combinatorial optimization Vol. 41; no. 2; pp. 318 - 327
• Main Authors: He, Zaobo, Sai, Akshita Maradapu Vera Venkata, Huang, Yan, seo, Daehee, Zhang, Hanzhou, Han, Qilong
• Format: Journal Article
• Language: English
• Published: New York Springer US 01.02.2021; Springer Nature B.V
• Subjects: Agglomeration; Aggregates; Algorithms; Budgets; Combinatorics; Convex and Discrete Geometry; Data management; Data sampling; Datasets; Feature selection; Mathematical Modeling and Industrial Mathematics; Mathematics; Mathematics and Statistics; Operations Research/Decision Theory; Optimization; Privacy; State-of-the-art reviews; Theory of Computation; Data aggregation; Differential privacy; Sampling; kd-tree
• ISSN: 1382-6905; 1573-2886
• DOI: 10.1007/s10878-020-00666-1

Privacy-preserving data aggregation is an important problem that has attracted extensive study. The state-of-the-art techniques for solving this problem is differential privacy, which offers a strong privacy guarantee without making strong assumptions about the attacker. However, existing solutions cannot effectively query data aggregation from high-dimensional datasets under differential privacy guarantee. Particularly, when the input dataset contains large number of dimensions, existing solutions must inject large scale of noise into returned aggregates. To address the above issue, this paper proposes an algorithm for querying differentially private approximate aggregates from high-dimensional datasets. Given a dataset D , our algorithm first develops a ε ′ -differentially private feature selection method that is based on a data sampling process over a kd-tree, which allows us to obtain a differentially private low-dimensional dataset with representative instances. After that, our algorithm samples independent samples from the kd-tree aiming at obtaining ( α ′ , δ ′ ) -approximate aggregates. Finally, a model is proposed to determine the relevance between privacy and utility budgets such that the final aggregate still satisfies the accuracy requirements specified by data consumers. Intuitively, the proposed algorithm circumvents the dilemma of both dimensionality and the height threshold of kd-tree, as it samples a low-dimensional dataset S and queries aggregates from S , instead of the kd-tree. Satisfying user-specified privacy and utility budgets after multiple-stages approximation is significantly challenging, and we presents a novel model to determine the parameters’ relevance.