Can Attackers Learn from Samples?

Abstract

Sampling is often used to achieve disclosure limitation for categorical and microarray datasets. The motivation is that while the public gets a snapshot of what is in the data, the entire data is not revealed and hence complete disclosure is prevented. However, the presence of prior knowledge is often overlooked in risk assessment. A sample plays an important role in risk analysis and can be used by a malicious user to construct prior knowledge of the domain. In this paper, we focus on formalizing the various kinds of prior knowledge an attacker can develop using samples and make the following contributions. We abstract various types of prior knowledge and define measures of quality which enables us to quantify how good the prior knowledge is with respect to the true knowledge given by the database. We propose a lightweight general purpose sampling framework with which a data owner can assess the impact of various sampling methods on the quality of prior knowledge. Finally, through a systematic set of experiments using real benchmark datasets, we study the effect of various sampling parameters on the quality of prior knowledge that is obtained from these samples. Such an analysis can help the data owner in making informed decisions about releasing samples to achieve disclosure limitation.