Evaluating adversarial partitions

Abstract

In this paper, we introduce a framework for measuring unlinkability both per subject and for an entire system. The framework enables the evaluator to attach different sensitivities to individual items in the system, and to specify the severity of different types of error that an adversary can make. These parameters, as well as a threshold that defines what constitutes a privacy breach, may be varied for each subject in the system; the framework respects and combines these potentially differing parametrisations. It also makes use of graphs in a way that results in intuitive feedback of different levels of detail. We exhibit the behaviour of our measures in two experimental settings, namely that of adversaries that output randomly chosen partitions, and that of adversaries that launch attacks of different effectiveness. © 2010 Springer-Verlag Berlin Heidelberg.