Paper: Utility vs Privacy (DP and e.g. Disclosure Risk)

[ots22]

Write a paper or technical report that compares differentially-private synthesis methods for their utility, as the privacy parameter (epsilon) varies.

• identify suitable synthesis methods (based on literature review of those offering DP)
• decide on a variety of input datasets covering the problems we are interested in, and corresponding parameters for them
• decide on the measures of utility (problem/dataset dependent)
• could also compute the Disclosure Risk score for the syntheses (since it depends only on input and output data)
• produce some figures of utility vs privacy (e.g. like those in the discussion on Experiments: utility vs privacy (as measured by disclosure risk) #120)
• capture any new methods and the workflow for the figures in QUIPP-pipeline, as well as the run inputs used to generate the results (for reproducibility and to allow extension to other methods, or refinement of existing methods)

[ots22]

Check this idea with a DP expert first, to make sure the comparisons we want to make are sensible.

[gmingas]

This working paper is related to what we are proposing here, focusing on DP-GANs only.

[ots22]

Good find!

[gmingas]

It was suggested by Giovanni Cherubin during the synthetic data meeting Kasra and I had earlier today.