Beschreibung

"What privacy guarantees can synthetic data satisfy even without formal guarantees during the training of the synthesizer? In this paper, we explore this question using synthesizers under simplified settings to show that the privacy guarantees offered by these synthesizers can be directly translated into a ρ-zCDP guarantee. We further explore the conditions under which this equivalence holds and show that it is significantly harder to get formal privacy guarantees for more realistic synthetic data models. Furthermore, we discuss under which conditions such synthetic data can be used to draw valid statistical inferences." (Author's abstract, IAB-Doku) ((en))