Skip to content

Inferred dataset properites need to be "sanitized" to ensure privacy safety w.r.t. source data #20

Description

@mmcdermott

Principally, this means we need to ensure:

  1. There are no direct join keys (e.g., timestamps, time-deltas, code names that correspond to a small number of patients, etc.). Code names are currently discarded so that point is not an issue remotely here, but it is worth noting in general.
  2. Continuous / non-categorical distributions are used where possible to ensure support is not concentrated on only real data.
  3. High-precision outputs are noised.

Note that none of these are truly serious concerns here, as there is no real data relating to medical codes or laboratory results preserved, and all aspects of the synthetic patients' record are generated independently (so it is literally impossible even with the current approach to infer anything about an existent real patient given a synthetic patient), but to avoid even the possibility of the appearance of a membership inference attack, it is best to be extra careful.

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    Inference of Generation ConfigsFor issues relating to the inference of data generation configs.bugSomething isn't workingpriority:criticalCritical priority; warrants an ASAP hotfix or new release candidate

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions