Secure k-Anonymization Linked with Differential Identifiability (Workshop)

Most k-anonymization mechanisms that have been developed presently are vulnerable to re-identification attacks, e.g., those generating a generalized value based on input databases. k-anonymization mechanisms do not properly capture the notion of hiding in a crowd, because they do not impose any constraints on the mechanisms. In this paper, we define \((k,\rho )\)-anonymization that achieves secure k-anonymization notion linked with differential identifiability under the condition of privacy parameter \(\rho \). Both differential identifiability and k-anonymization limit the probability that an individual is re-identified in a database after an adversary observes the output results of the database. Furthermore, differential identifiability can provide the same strong privacy guarantees as differential privacy. It can make k-anonymization perform securely, while \((k,\rho )\)-anonymization achieves the relaxation of the notion of differential identifiability, which can avoid a lot of noise and help obtain better utility for certain tasks. We also prove the properties \((k,\rho )\)-anonymization under composition that can be used for application in data publishing and data mining.