Reviewing re-identification risk

SkillDatabases & data

Lets your agent score how likely a de-identified dataset is to be re-identified and write up the risk findings.

Available today. Use it from your connected AI after setup.

Connect ahel once, and every AI you use reads what you have installed.

Then ask your AI: use the Reviewing re-identification risk skill

About this capability

Run expert-determination-style quasi-identifier risk scoring (k-anonymity, l-diversity) plus OpenMed's empirical re-identification attack on a de-identified dataset, then document residual risk in a defensible memo. Use when the user needs HIPAA Expert Determination (45 CFR 164.514(b)(1)) support, a

What this skill tells your AI

The instructions your AI receives, as published by maziyarpanahi/openmed in skills/reviewing-reidentification-risk/SKILL.md and read by ahel’s review.

Removing direct identifiers is not enough. A record stripped of name, SSN, and MRN can still be singled out by a combination of quasi-identifiers — age, ZIP/region, admission date, sex, rare diagnosis. The HIPAA Expert Determination pathway (45 CFR 164.514(b)(1)) requires a qualified person to apply statistical methods and document that the risk of re-identification is "very small." This skill produces that evidence: quasi-identifier risk metrics (k-anonymity, l-diversity) plus OpenMed's empirical re-identification attack, written up as a residual-risk memo.

When to use

  • After direct-identifier removal passes auditing-deid-leakage (no leaks) and you must decide whether the dataset is releasable.
  • The user invokes Expert Determination, asks for a re-identification risk score, k-anonymity, l-diversity, or a "very small risk" determination memo.
  • You need an adversarial linkage attack — modeling an attacker with auxiliary data — not just a structural metric.

Quick start

from openmed.eval.attacks.reid import run_reid_attack, run_reid_benchmark

# Synthetic de-identified records; each row is the released, de-id'd data.
deidentified = [
    {"record_id": "r1", "text": "[NAME], 47F, ZIP 021xx, admitted 2024-03."},
    {"record_id": "r2", "text": "[NAME], 47F, ZIP 021xx, admitted 2024-03."},
    {"record_id": "r3", "text": "[NAME], 88M, ZIP 597xx, admitted 2024-03."},  # singleton
]
# Auxiliary = what an attacker might already hold (e.g. a voter list).
auxiliary = [{"record_id": "v9", "text": "88M ZIP 597xx"}]

result = run_reid_attack(
    fixtures=[],                          # bring your own records below
    deidentified_records=deidentified,
    auxiliary_records=auxiliary,
)
metric = result.to_metric()
print(metric["aux_linkage_rate"],        # empirical linkage success
      metric["k_min"],                   # smallest equivalence-class size
      metric["singleton_count"],         # k=1 records (uniquely identifiable)
      metric["quasi_identifier_count"])

k_min is the population k-anonymity floor across the dataset; a k_min of 1 means at least one record is unique on its quasi-identifiers and is the highest re-identification risk. aux_linkage_rate is the empirical attack: how often the adversary's auxiliary data successfully links back to a released record.

To run against the bundled golden suite and emit a leaderboard-style report:

report = run_reid_benchmark(
    suite="golden",
    deidentified_records=deidentified,
    auxiliary_records=auxiliary,
    output_markdown="reid_risk.md",
)

Workflow

  1. Enumerate quasi-identifiers (QIs). List every field an outsider could plausibly know and cross-reference: age/DOB, ZIP/region, dates of service, sex, race, rare conditions, provider. Direct identifiers should already be gone (verified by auditing-deid-leakage); QIs are what's left to worry about.
  2. Compute k-anonymity. For each equivalence class (records sharing the same QI combination), the class size is k. run_reid_attack returns k_min and the list of singleton_records (k=1). A common Expert Determination target is k ≥ a documented threshold (e.g. k ≥ 5 or k ≥ 11) for every record.
  3. Check l-diversity on sensitive attributes. k-anonymity hides which record, but if every record in a class shares the same sensitive value (e.g. all HIV-positive), the attribute leaks anyway. Require ≥ l distinct sensitive values per class; flag homogeneous classes.
  4. Run the empirical attack. run_reid_attack / run_reid_benchmark model an adversary with auxiliary_records and measure actual linkage success (aux_linkage_rate), residual leakage (leakage_rate), surrogate-consistency leaks, and date-shift-inversion leaks. Structural metrics bound risk; the attack demonstrates it.
  5. Generalize or suppress, then re-score. For singletons / low-k classes, coarsen QIs (age → age band, ZIP5 → ZIP3, exact date → month/quarter) or suppress the record, then re-run until k_min and linkage rate meet your documented threshold.
  6. Write the determination memo. Record the QIs considered, methods applied, k_min, l-diversity, the attack's aux_linkage_rate, the assumptions about attacker capability, and the conclusion that residual risk is "very small." Cite the metrics — never paste raw records into the memo.

Hand-off to / from OpenMed

  • From auditing-deid-leakage: only score QI risk once direct-identifier leakage is zero. A leak short-circuits the whole determination.
  • OpenMed calls: from openmed.eval.attacks.reid import run_reid_attack, run_reid_benchmark, generate_reid_leaderboard. The attack delegates to openmed.risk.risk_report for k-anonymity / linkage internals.
  • To evaluating-with-leakage-gates: register reid_leakage_rate as a gate in the eval harness so re-identification risk regressions fail CI.
  • From pseudonymizing-for-gdpr: pseudonymized output is still re-identifiable via QIs — run this attack before claiming a dataset is low-risk or anonymized.

Edge cases & gotchas

  • Expert Determination is a human judgment. OpenMed produces the statistics; a qualified expert signs the determination. The tool supports the memo, it is not the memo.
  • Auxiliary data assumptions drive the result. Linkage rate is only as meaningful as the auxiliary_records you model. Document the assumed attacker (motivated insider vs. public voter list) — different aux sets, different risk.
  • Singletons are the headline. A single k=1 record can sink a release; check singleton_count and singleton_records first.
  • Date-shift can be inverted. Preserving intervals across a date shift lets an attacker re-anchor the timeline; the attack flags date_shift_inversion_rate. Watch it when de-id used method="shift_dates".
  • No raw records in artifacts. Reports carry counts, rates, and offsets. Keep the underlying dataset out of the memo and out of logs.
  • Local-first. Run the attack on-device; never ship candidate-release data to a third party to "test" re-identifiability.

Standards & references

Signals

GitHub stars
5k
Forks
668
Last commit
Sep 2026
Advanced
Catalog kind
skill
Gateway key
reviewing-reidentification-risk
Source
github.com/maziyarpanahi/openmed