barissozudogru/piiscope-benchmark

Piiscope Structured PII Pattern Benchmark

A deterministic, privacy-safe benchmark for structured personal-data detectors. Every value is synthetic, reserved for documentation, or a published test credential. The dataset contains no records collected from people, no customer data, and no transactable financial identifiers.

The benchmark is maintained with Piiscope, a local PII scanner and privacy-risk CLI. It can also evaluate compatible rule-based detectors that return one or more rule identifiers for a text field.

The focused Piiscope tool page collects the install path, operating limits, reproducible workflow, and source links in one place.

Dataset structure

The test split contains 512 rows built from 32 base cases and 16 English, German, Turkish, and Portuguese contexts.

• Positive cases cover emails, checksum-valid test cards and IBANs, fictional phone ranges, documentation IP addresses, URLs, dates, medical record markers, VAT markers, SWIFT markers, and synthetic passport patterns.
• Hard negatives cover malformed checksums, invalid network addresses, lowercase near misses for case-sensitive rules, short identifiers, version numbers, and commit-like hexadecimal strings.
• expected_rules is a multi-label detector contract.
• safety_basis explains why the embedded value does not represent a person or live account.

Reproduce the data

python benchmarks/piiscope-benchmark/generate.py

Generation uses only the Python standard library and contains no random step. The same source always produces the same ordered JSONL file.

Evaluate Piiscope

python -m pip install piiscope
python benchmarks/piiscope-benchmark/evaluate.py

The checked-in result records the exact Piiscope version and reports micro precision, micro recall, exact multi-label match rate, and hard-negative pass rate. Dictionary detectors are disabled so this first release measures the structured-pattern layer in isolation.

Published results

Piiscope	Micro precision	Micro recall	Micro F1	Exact match	Hard-negative pass
1.2.0	0.7941	0.9643	0.8710	0.7500	0.7143
1.3.0	1.0000	1.0000	1.0000	1.0000	1.0000

The 1.2.0 baseline was measured from the published PyPI wheel. The 1.3.0 row is the release result after the benchmark exposed phone-boundary, partial-IPv6, and cross-detector false positives. Full failures and per-rule counts are stored under results/.

Intended use

• Regression testing for deterministic structured PII detectors
• Comparing checksum and near-miss handling across releases
• Demonstrating privacy-scanner behavior without publishing real PII
• Teaching safe benchmark construction for privacy tooling

Limitations

• This is a pattern-detection benchmark, not a model of real-world PII prevalence.
• It does not measure demographic fairness or multilingual named-entity recall.
• It deliberately excludes positive national identity numbers when no clearly reserved test range is available.
• Repeated synthetic values test context handling, not memorization resistance.
• A high score does not prove that a scanner is sufficient for compliance.

Licensing and citation

The dataset is released under CC BY 4.0. Generator and evaluation scripts in the Piiscope repository remain under Apache 2.0.

When citing the benchmark, link both the dataset and the evaluated Piiscope release so the detector behavior can be reproduced.