Build "plausible deniability" privacy metric

[ots22]

(See further discussion in #60)

Implement (as a "privacy metric" in the pipeline), the Plausible Deniability metric
(code here)

• include the implementation directly as both a method and a privacy measure
• customize for any method that depends on a seed (e.g. synthpop - Include synthpop as a method in the pipeline #81)

[ots22]

See alan-turing-institute/QUIPP-pipeline#25

Context: the implementation above needs a few steps, and some additional metadata that can be computed from the input data (which the implementation doesn't do).

Steps:

• make example parameter json file for the method
• script to generate additional metadata from input data and in the format required by sgf
  • create configuration files (from parameter json and data) (this is "my.cfg" in the sgf example)
  • split data ("stats" - training, "records" - generating)
  • "attrs" - set of values in a column
  • "grps" - bins: treat as binwidth of 1, for now (this will be the same as "attrs", without the label)
  • DAG - function to compute thresholded covariance matrix, function to compute merit score from this (see paper). DAG format needs: edge heads in vertex order, separate traversal order (must be topological order). See README.pdf
• Write "run" for the method (for Makefile/pipeline)

[ots22]

Document this!

[ots22]

Still to do:

• document what we have so far
• clearly separate privacy output step from synthesis output step
  • privacy metric (in this case) will transform the privacy parameters (k, eps_0, gamma) in the input config json file to (eps, delta) differential privacy parameters in the output.

[ots22]

Could be useful to refer to the current work on https://github.com/alan-turing-institute/QUIPP-pipeline/tree/feature/82-disclosure-risk

[ots22]

PR alan-turing-institute/QUIPP-pipeline#27 now merged