Skip to content

Describe per-record DP in the County Business Patterns record (#151) - #224

Open
autoCAT6 wants to merge 1 commit into
opendp:mainfrom
autoCAT6:cbp-per-record-dp
Open

autoCAT6 wants to merge 1 commit into
opendp:mainfrom
autoCAT6:cbp-per-record-dp

Conversation

@autoCAT6

Copy link
Copy Markdown

Refs #151.

The CBP Differential Privacy Demonstration Tables (Round 1) use per-record DP in stage one. Establishments are split into pieces before a zCDP mechanism runs, so an establishment's guarantee depends on how many pieces it is split into. The record described this as plain zCDP with a business establishment as the privacy unit and rho = 12.058. This PR keeps those values and uses the details fields to say what they mean.

Changes to deployments/county_business_patterns.yaml:

  • variant_name_details (new): cites the P-per-record zCDP definition in the VLDB paper (Definition 6 and Lemma 1, p. 3140), and the Census Bureau's description of stage one as establishment splitting followed by a DP mechanism (slides 29-30).
  • privacy_unit_details: keeps the existing quote. Adds that adjacency is adding or removing one original establishment, and that split establishments get a weaker guarantee. The slides give that bound in pure DP terms, |A(r)| x epsilon (slide 30). The paper gives it in zCDP terms, |A(r)|^2 x rho (Lemma 8, p. 3142). Also lists the thetas from slides 50-51 and notes that the tables were computed from synthetic microdata.
  • privacy_parameters_details: keeps the existing quote. Breaks 12.058 into the four per-measure budgets (slides 50-51), and says it is a total across measures. The sources checked do not give a common splitting factor that would turn it into a bound for each establishment. Says the second stage is "Not formally private" (slide 46).
  • Evidence sources: about now points to a February 2025 Wayback copy, because the current page no longer has the quoted paragraph. Added the webinar slides, the Round 1 demonstration tables page and the paper. Added a comment to webinar saying the page now marks this information as no longer current.

Two points are inferences, and the record labels them that way:

  1. theta as the splitting threshold. The slides list a theta for each measure except the establishment count, and slide 47 shows an example "PRDP with theta = 10 and rho = 0.1". None of the Census Bureau materials I found defines theta directly.
  2. rho = 12.058 for an establishment that is not split under any measure. This follows from zCDP composition. The Census Bureau does not state it.

Questions for maintainers:

  • I kept variant_name: Zero-concentrated DP and privacy_unit: Business establishment and put the caveats in the details fields, so the schema stays as it is. If you would rather use Custom for the variant, or describe the unit differently, I can change it.
  • The issue asks whether the splitting thresholds were ever published. The webinar slides list a theta for every measure except the establishment count, so I added them, with the inference noted above.

Locally on Python 3.14, pre-commit run --all-files, python scripts/check.py --skip check_urls and pytest -vv pass, and check_urls passes for this file.

The CBP demonstration tables protect stage one with per-record DP:
establishments are split before a zCDP mechanism runs, so the level
of protection depends on how many pieces an establishment is split
into. The record described this as plain zCDP over establishments
with rho = 12.058.

Keep the existing enum values and the rho value, and explain in the
details fields what they mean: the per-record definition from the
VLDB paper, the split-dependent bound as the slides (pure DP) and the
paper (zCDP) each state it, the per-measure budgets and thetas from
the webinar slides, and the condition under which 12.058 applies.

Point the about citation to a February 2025 archived copy, since the
current page no longer has the quoted paragraph. Add the webinar
slides, the Round 1 demonstration tables page and the paper as
evidence sources, and note that the webinar page now marks this
information as no longer current.

Refs opendp#151
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant