Disclosure risk metric planning / tracking

[jhseeman]

Disclosure risk metrics planning

This issue will be used to plan updates for disclosure risk metrics in syntheval

Confidential data baseline assessments

• [ ] Methods for identifying existing confidential records with high disclosure risk
  • [ ] disc_baseline_k_anon(conf_data): row-wise $k$-anonymous values
  • [ ] disc_baseline_t_closeness(conf_data): row-wise $t$-closeness values
  • [ ] disc_baseline_l_diversity(conf_data): row-wise $l$-diversity values
• [ ] Methods for identifying arbitrary records worth evaluating in holdouts
  • [ ] disc_baseline_lra(conf_tables): linear reconstruction attack from a collection of count tables (link)
  • [ ] disc_baseline_make_canaries(conf_data): create artificial high-risk records for holdout data (e.g., "canaries" (link)

Membership inferences from synthetic data

• [ ] Membership inference updates for arbitrarily holdouts (link)
  • [ ] disc_mit(...) updates for multiple synthetic data replicates
  • [ ] disc_mit(...) updates for disaggregated records
  • [ ] disc_mit(...) updates for mechanism adaptivity
• [ ] Linkage attacks
  • [ ] disc_linkage_recon(synth_data, recon): Linkage attack from synthetic data and partial reconstruction

Attribute inferences

• [ ] disc_ait(synth_data, test_records): attribute inference for test_records using synthetic data-based models
• [ ] disc_ait_compare(synth_data, test_records, holdout_data): attribute inference for test_records comparing differences between using synthetic and holdout data (link)

[awunderground]

Confidential data baseline assessments

I would appreciate functionality/best practices for working with continuous variables and mixed-type data.

Membership inferences from synthetic data

Can you share a little more detail about the linkage attack functionality?

1. I can imagine major differences between methods for partially and fully synthetic data.
2. What is the direction of the linkage?

Attribute inferences

I've done some crude work on this. Let me know how I can help! The discriminator workflow we added is pretty flexible and leverages library(tidymodels).