Rare-case privacy stressors¶

Average synthetic populations seldom contain the unusual attribute combinations that make a record re-identifiable: a centenarian in a hamlet of a few hundred people, a lighthouse keeper with a rare diagnosis, one of quadruplets with a 29 February birth date. generate_rare_case_stress_set builds those combinations on purpose, from a seed, and labels every record with the direct identifiers it carries and with how much each quasi-identifier contributes to its re-identification risk. Privacy attack suites can then be scored against known ground truth.

from openmed.training.synthetic.rare_case_stress import (
    RareCaseStressConfig,
    generate_rare_case_stress_set,
)

stress_set = generate_rare_case_stress_set(
    30,
    seed=7,
    config=RareCaseStressConfig(background_classes=10),
)
print(stress_set.summary()["stress_singleton_count"])
record = stress_set.records[0]
print(record.profile_id, record.equivalence_class_size, record.dominant_dimension)

Generation is deterministic: the same seed and configuration always yield the same records, note text, offsets, and dataset digest. It uses a local random.Random, reads no files, writes no logs, and makes no network call.

Dimensions and tiers¶

Each dimension draws from a small, repository-authored vocabulary split into common, uncommon, and rare tiers. Every value carries a synthetic frequency: an order-of-magnitude prior, not a census, registry, or prevalence statistic. Place names are fictional.

Dates are released at a precision. Their frequency is the share of people of the same released age who share the released value, so the birth year adds no information beyond age and age and date are not double counted. Leap-day birth dates are always consistent with the released age at the configured reference_date (default 2025-06-30); a configuration that cannot place a 29 February birth date inside the drawn age band raises ValueError.

rare_case_vocabulary() returns every opaque code (for example occupation.rare.04) with its tier and frequency, without the surface values.

Configuring combinations¶

• RareCaseProfile(rare=(...), uncommon=(...)) requests a tier per dimension; unnamed dimensions stay common. RareCaseProfile.from_tiers({...}) accepts a {dimension: tier} mapping, and profile_id gives a stable, value-free name such as rare:occupation;uncommon:age.
• default_rare_case_profiles(order) returns every combination of order simultaneously rare dimensions. The default configuration uses the 15 pairwise combinations.
• RareCaseStressConfig.profiles is cycled in order across stress records, so every configured combination is covered once count reaches the number of profiles.
• background_classes and background_class_size add all-common crowd records that share one quasi-identifier tuple per class, so stress records are rare relative to a k-anonymous background.
• include_direct_identifiers=False omits the synthetic name and medical record number.

Risk labels¶

record.risk_labels() is value-free. For each attribute it reports the dimension, column, canonical label, identifier class, re-identification category, tier, opaque code, character offsets, and:

• frequency and surprisal_bits (-log2(frequency)), the information the released value carries;
• contribution_share, the attribute's share of the record's combined surprisal;
• risk_level (low, medium, high for common, uncommon, rare);
• safe_harbor_flags, descriptive flags such as age_over_89, date_indicates_age_over_89, date_more_specific_than_year, and small_geographic_unit (20,000 people or fewer);
• sensitive_attribute for diagnoses and relative_role for family relationships, using the roles from family-history attribution.

Record-level labels add equivalence_class_size (the dataset k computed over the generated set), dataset_singleton, combined_surprisal_bits, expected_population_matches against a fictional reference population of ten million, a Poisson population_unique_probability, and the dominant_dimension. The combined figures assume independent attributes, a deliberate simplification.

Identifier classes follow the core label policy: the synthetic name, record number, and a full birth date are DIRECT_IDENTIFIER. Every other dimension is scored as a QUASI_IDENTIFIER, including diagnoses that core label metadata treats as clinical concepts, because rare conditions enable linkage. Categories reuse the re-identification profiler vocabulary (geography, age, date, rare_condition) and add occupation and family_relationship.

Using the set with privacy attack suites¶

rows(include_direct_identifiers=False) removes the synthetic name and medical record number. Full birth dates remain in the linkage key even though the core label policy classifies them as direct identifiers. quasi_identifier_fields names that complete linkage key; this export is an attack fixture, not a de-identified release.

from openmed.eval.attacks import linkage_attack
from openmed.risk.reid import risk_report

released = stress_set.rows(include_direct_identifiers=False)
fields = stress_set.quasi_identifier_fields

report = risk_report(released, quasi_identifier_fields=fields)
attack = linkage_attack(released, stress_set.rows(), quasi_identifiers=fields)

truth = {r.record_id for r in stress_set.records if r.is_dataset_singleton}
assert {item["record_id"] for item in report["singleton_records"]} == truth
assert attack.unique_matches == len(truth)

to_dicts() returns training records with note text, offset-aligned spans in the shape OpenMed risk profilers read, typed fields, the value-free risk labels, and synthetic provenance metadata.

Privacy boundary¶

• Every value is synthetic. No restricted dataset, registry, or real record is read or bundled.
• summary(), risk_labels(), rare_case_vocabulary(), configuration summaries, and object representations contain codes, tiers, offsets, counts, and sha256: digests only. Generated values appear only in the records' text, fields, and rows.
• Validation errors never echo the rejected value.
• The labels describe synthetic risk for testing attack suites and de-identification pipelines. They are not an Expert Determination, a compliance certification, or a clinical statement.