On Privacy and Accuracy in Data Releases (Invited Paper)

In this paper we study the relationship between privacy and accuracy in the context of correlated datasets. We use a model of quantitative information flow to describe the the trade-off between privacy of individuals\u27 data and and the utility of queries to that data by modelling the effectiveness of adversaries attempting to make inferences after a data release. We show that, where correlations exist in datasets, it is not possible to implement optimal noise-adding mechanisms that give the best possible accuracy or the best possible privacy in all situations. Finally we illustrate the trade-off between accuracy and privacy for local and oblivious differentially private mechanisms in terms of inference attacks on medium-scale datasets