Abstract by Mitchell Wassom
Using Multiway Layouts to Rethink Differential Privacy
The concern over privacy has limited the sharing of data. In recent years methods of proposed data obfuscation allow users to share partial information in data sets. One index of the level of privacy in data sharing is “differential privacy.” A downfall to current approaches to this privacy is that, to make datasets secure, people add excessive random noise, lowering its interpretability.
This paper discusses a different approach – discussing differential privacy in the context of the multiway layout for categorical data. By considering only partially saturated models of cell counts this paper illustrates how differential privacy is obtained by rounding without having input excessive random noise into the data table. The procedure is also extendable to continuous data. We use a large health data set as an example.