pii-redactor-au 0.4.2

Pre-ingestion PII de-identification for Australian government data, using zero-shot LLM detection with checksum validation.

PII Redactor

Pre-ingestion PII de-identification for Australian government data.

Implements the methodology from Wiest et al., Deidentifying Medical Documents with Local, Privacy-Preserving Large Language Models (NEJM AI, 2024), extended with Australian Commonwealth identifier detection and checksum validation.

v0.4.1 — AU organisation + location recognisers (2026-05-20). Per-sector sensitivity lift +0.64 to +6.76 pp across federal/state-health/legal/medtech benches; organisation per-category recall +13-56 pp; location +33 pp on federal. Frozen Wiest bench held exactly (Medical 99.71%/0 leaks, Gretel 93.67%/2 leaks). The companion LoRA adapter JimmyBhoy/redact-au-1b ships on HF Hub (apache-2.0) for the two-tier AU-government + clinical deployment path. See CHANGELOG.md for the full v0.4.1 changeset.

What it does

Sits between a data source and an ingestion target in your pipeline. Documents come in containing PII; documents go out with PII replaced by structured placeholders. Original values live only in an encrypted audit log, recoverable by audit ID with the right key.

   Source → [PII Redactor] → Ingestion target (warehouse, search, model training, ...)
                ↓
        encrypted audit log

Why this design

Four things make this approach work for production AU government + clinical workloads:

1. LoRA-tuned token classifier on openai/privacy-filter (the default in v0.4.x). 1.5B-param / 50M-active MoE token classifier, fine-tuned on a ~55k-doc AU-synthetic corpus covering all 11 AU regulatory categories. Wiest-equivalent recall on AU clinical narratives with 5-6 docs/sec inference on a single RTX 4090. The LoRA adapter is published separately at JimmyBhoy/redact-au-1b (Apache 2.0).
2. Two-tier product. Default tier (Tier 1) drops the slow narrative-NER pass for ≥3 d/s throughput on AU compliance workloads. Opt-in Tier 2 (PIIR_LLAMA_BACKEND=vllm) layers a Llama-3.1-8B narrative pass on top for general-PII workloads at lower throughput but higher recall on non-AU formats.
3. Local inference. No data leaves your boundary. Important for OFFICIAL / OFFICIAL: Sensitive deployments. PROTECTED requires IRAP (see grants/IRAP-scoping.md in the companion redact-au repo for the assessment path).
4. Australian checksum validation (our addition). The token classifier is tuned for recall; downstream validators add precision. A nine-digit number that looks like a TFN but fails the mod-11 weighted sum gets dropped. Same for ABN (mod-89), ACN, Medicare. This filters most false positives on numeric identifiers.

Detected categories

Category	Validation	Notes
name	none	Personal names of individuals
address	none	Postal or street addresses
patient_id	context-bound regex + LLM	Labelled patient identifiers
medical_record_number	context-bound regex + LLM	MRN, URN, hospital number
healthcare_identifier	context-bound regex + LLM	IHI or labelled healthcare identifier
date_of_birth	none	Distinguished from generic dates by context
date	none	Other dates
phone	regex format	AU mobile and landline formats
email	regex format	RFC 5322 simplified
tfn	mod-11 checksum	8 or 9 digits, ATO algorithm
medicare	checksum	10 or 11 digits, includes IRN handling
abn	mod-89 checksum	11 digits, ATO algorithm
acn	checksum	9 digits, ASIC algorithm
driver_licence	none	State-specific formats
passport	none	Australian passport pattern
bsb_account	format	6-digit BSB + account
centrelink_crn	none	Customer Reference Number
organisation	gazetteer + suffix pattern	Companies (Pty Ltd, Limited), gov agencies, hospitals, universities, banks, law firms (v0.4.1+)
location	gazetteer + state-abbrev + postcode	Suburbs, states, postcodes, regions — standalone references (v0.4.1+)
url, ip_address	regex	Standard formats

Validators that exist will drop invalid detections from the output. The validation result is preserved on each span so downstream consumers can see what was confirmed vs. flagged.

Measured performance (2026-05-20)

v0.4.1 frozen bench — RTX 4090, concurrent×8 HTTP:

Fixture	Tier 1 (default, LoRA, llama off)	Tier 2 (opt-in, +llama vLLM)	Legacy CPU baseline (v0.1.2 Ollama)
Gretel-100 sensitivity	93.67% (2 leaks)	97.89% (0 leaks)	99.58% (1 leak)
Medical-50 sensitivity	99.71% (0 leaks)	100.00% (0 leaks)	100.00% (0 leaks)
Throughput (docs/sec)	5-6	1.4	0.07
Mean latency / doc	~150-300 ms	~700 ms	13.8-20.4 s
Hardware	RTX 4090 (24 GB)	RTX 4090 + vLLM	Ryzen 7 7840HS CPU

All three tiers clear the Wiest et al. 99.4% sensitivity bar on the medical fixture (Wiest target = 99.4%; we hit 99.71% / 100% / 100%). The Gretel-100 fixture covers global+social PII broader than our AU-government wedge — Tier 2 closes that gap when the workload includes international phone formats, social handles, or foreign locality names.

Multi-sector v0.4.1 head-to-head (920 docs across federal-gov / state-health / legal / medtech, vs Microsoft Presidio):

Sector	redact-au v0.4.1 (Tier 1)	Microsoft Presidio + AU regex	Δ
federal-gov-official	93.15%	78.32%	+14.83 pp
state-health-hospitals	93.55%	86.05%	+7.50 pp
legal-small-mid	85.75%	69.92%	+15.83 pp
medtech-health-ai	91.57%	77.72%	+13.85 pp

100% recall on AU regulatory IDs (TFN/ABN/ACN/Medicare/IHI/MRN/BSB/Centrelink-CRN) with checksum validation across all four sectors. Per-category breakdown + leak taxonomy in the companion redact-au repo's benchmarks/ directory.

Per-leak diagnostics, severity grading, and full per-category recall are in CHANGELOG.md (v0.4.0 + v0.4.1) and the v0.0.x → v0.4.1 history under .planning/.

Three integration paths

1. Python library (recommended for in-process pipelines)

# From PyPI
pip install pii-redactor-au

# Or editable from source
pip install -e .

The PyPI distribution name is pii-redactor-au (the unqualified pii-redactor is owned by another project). The import name remains pii_redactor — no code change for consumers.

from pii_redactor import build_pipeline

pipeline = build_pipeline()  # reads config from env
result = pipeline.process_document(text)

print(result.redacted_text)
print(f"Found {result.pii_count} PII items")
print(result.pii_table())  # safe extracted-PII table, no original values
print(f"Audit ID: {result.audit_id}")

2. FastAPI service (for cross-language pipelines)

docker compose -f docker/docker-compose.yml up -d
curl localhost:8000/health

curl -X POST localhost:8000/redact \
  -H "X-API-Key: $PIIR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"text": "TFN: 123 456 782, contact: [e-mail]"}'

Endpoints:

• POST /redact — single document
• POST /redact/batch — list of documents
• POST /reidentify — recover original values from audit_id (requires audit key)
• GET /health, GET /info

3. Gradio Space (for demo and stakeholder review)

python app.py
# or push to HF Spaces with HF_TOKEN set as a Space secret

The Space uses HF Inference API as the backend so it works without llama.cpp. Same library, thin UI wrapper.

Configuration

All configuration via environment variables prefixed PIIR_. See .env.example for the full list. Key ones:

Variable	Default	Purpose
PIIR_BACKEND	transformers_au_finetuned	mock | llama_cpp | hf | ollama | transformers_au | transformers_au_finetuned
PIIR_LORA_ADAPTER	JimmyBhoy/redact-au-1b	HF Hub repo or local path to LoRA adapter (only for transformers_au_finetuned)
PIIR_LLAMA_BACKEND	disabled	disabled | vllm | ollama | auto — Tier 2 narrative-NER pass
PIIR_LLAMA_GATE	always	always | confidence | never — when Tier 2 is enabled, gate per-doc
PIIR_VLLM_URL	http://localhost:11500	vLLM OpenAI-compatible endpoint
PIIR_OLLAMA_URL	http://localhost:11434	Ollama server endpoint
PIIR_LLAMA_CPP_URL	http://localhost:8080	llama.cpp server endpoint
PIIR_HF_TOKEN	—	HF Inference API token (for hf backend)
PIIR_REGEX_USERNAME	false	Toggle username regex (off by default — false positives risk)
PIIR_REGEX_ORGANISATION	true	Toggle organisation gazetteer + suffix recogniser (v0.4.1+)
PIIR_REGEX_LOCATION	true	Toggle location gazetteer + state/postcode recogniser (v0.4.1+)
PIIR_API_KEY	—	FastAPI auth (unset = open)
PIIR_AUDIT_KEY	—	Fernet key for audit encryption
PIIR_PLACEHOLDER_STYLE	numbered	numbered | category | asterisk

Generate keys:

# API key (for X-API-Key header)
python -c "import secrets; print(secrets.token_urlsafe(32))"

# Audit encryption key (Fernet)
python -c "from cryptography.fernet import Fernet; print(Fernet.generate_key().decode())"

Audit and re-identification

Every detection is written to an encrypted JSONL audit log:

{
  "audit_id": "uuid",
  "document_id": "uuid",
  "timestamp": "...",
  "category": "tfn",
  "span_start": 14,
  "span_end": 25,
  "value_encrypted": "gAAAAAB...",
  "placeholder": "[REDACTED_TFN_001]",
  "confidence": 1.0,
  "validator_passed": true,
  "model_used": "llama-3-8b-instruct"
}

With the audit key, an authorised caller can recover originals:

originals = pipeline.audit.reidentify(audit_id)
# → list of dicts including the decrypted `value`

If no audit key is configured, the log records that PII was detected and what category, but not the values. We refuse to write plaintext PII to the audit log under any circumstances. This is a deliberate fail-safe.

Privacy Act 1988 alignment

Three properties of this design that map to APP obligations:

• APP 11 (security). Original values are encrypted at rest in the audit log. AES-128-CBC + HMAC via Fernet. Re-identification requires the key, which is held outside the redactor.
• APP 11.2 (destruction). Setting PIIR_AUDIT_ENABLED=false disables logging entirely. Spans returned downstream never carry original values regardless of audit configuration.
• APP 12 (access and correction). The audit_id returned with every redaction lets a data subject's request be linked back to specific source documents.

This is not a substitute for a legal review. It's infrastructure that makes APP-aligned operation tractable.

Testing

pytest tests/ -v

Current state: 137 tests, all passing (v0.4.1). Validators exercised against published test values from ATO (TFN, ABN), ASIC (ACN), and synthetically valid Medicare numbers. Pipeline tests use the MockClient backend (regex-only) so they run offline without an LLM. Regression coverage spans every release: v0.1.x phone/BSB/ABN edge cases, v0.2.x hybrid pipeline integration, v0.3.x vLLM + concurrency, v0.4.x LoRA backend + org/location recognisers + llama-disabled default + fallback-chain.

Architecture

pii_redactor/
├── models.py                # PIICategory (+ ORGANISATION, LOCATION since v0.4.1), PIISpan, RedactionResult
├── validators.py            # AU checksum algorithms (TFN, ABN, ACN, Medicare)
├── prompts.py               # Zero-shot prompt templates (legacy llama path)
├── grammar.py               # GBNF grammar for llama.cpp constrained sampling
├── llm_client.py            # Backends: LlamaCpp, HFInference, Ollama, Mock
├── detector.py              # chunk → prompt → parse → validate → merge
├── redactor.py              # Apply placeholders, three styles
├── audit.py                 # Fernet-encrypted JSONL audit
├── pipeline.py              # Public API: build_pipeline(), process_document()
├── config.py                # env-driven Pydantic settings
├── data/gazetteers/         # AU org + suburb + postcode + state gazetteers (v0.4.1+)
└── hybrid/                  # Hybrid pipeline (v0.2.x+)
    ├── openai_backend.py    # openai/privacy-filter token-classifier wrapper
    ├── finetuned_backend.py # PEFT-loaded LoRA adapter on openai/privacy-filter (v0.4.x+)
    ├── llama_pass.py        # Ollama Llama-3.1-8B narrative pass (Tier 2)
    ├── vllm_pass.py         # vLLM AWQ-Marlin narrative pass (Tier 2, RTX)
    ├── au_resolver.py       # Checksum + label-prior resolution for AU IDs
    ├── au_org_loc.py        # Organisation + location recognisers (v0.4.1+)
    ├── regex_supplement.py  # Username + AU phone/address fallback
    └── pipeline.py          # HybridDetector orchestrator + tier selection

api/main.py                  # FastAPI HTTP service
app.py                       # Gradio Space entry point
tests/                       # Pytest suite (137 tests as of v0.4.1)
docker/                      # Dockerfile + compose stack

Source

Wiest IC, Leßmann ME, Wolf F, et al. Deidentifying Medical Documents with Local, Privacy-Preserving Large Language Models: The LLM-Anonymizer. NEJM AI 2024. DOI: 10.1056/AIdbp2400537

License

MIT

Local project consolidation

This project is now consolidated under C:\Users\j_car\KnowledgeGraph\tools\pii-redactor.

Related reference material is kept in docs/references/, including the NEJM AI LLM-Anonymizer paper used to ground the de-identification design.