Generalization-Based Privacy-Preserving Data Collection

Abstract

In privacy-preserving data mining, there is a need to consider on-line data collection applications in a client-server-to-user (CS2U) model, in which a trusted server can help clients create and disseminate anonymous data. Existing privacy-preserving data publishing (PPDP) and privacy-preserving data collection (PPDC) methods do not sufficiently address the needs of these applications. In this paper, we present a novel PPDC method that lets respondents (clients) use generalization to create anonymous data in the CS2U model. Generalization is widely used for PPDP but has not been used for PPDC. We propose a new probabilistic privacy measure to model a distribution attack and use it to define the respondent’s problem (RP) for finding an optimal anonymous tuple. We show that RP is NP-hard and present a heuristic algorithm for it. Our method is compared with a number of existing PPDC and PPDP methods in experiments based on two UCI datasets and two utility measures. Preliminary results show that our method can better protect against the distribution attack and provide good balance between privacy and data utility.