Privplay

PHI/PII Classification Engine with Training Pipeline.

Detection Stack (Defense-in-Depth)

Privplay uses multiple detection layers for maximum HIPAA coverage:

┌─────────────┐   ┌─────────────┐   ┌─────────────┐   ┌─────────────┐
│ Transformer │   │  Presidio   │   │    Rules    │   │    LLM      │
│ (Stanford)  │ → │  + Dicts    │ → │   (Regex)   │ → │  Verifier   │
│ PHI-specific│   │  PII + Med  │   │ HIPAA 18 IDs│   │  Uncertain  │
└─────────────┘   └─────────────┘   └─────────────┘   └─────────────┘

HIPAA 18 Identifier Coverage

#	Identifier	Detection Method
1	Names	Transformer + Presidio NER
2	Geographic	Presidio + Rules
3	Dates	Transformer + Presidio + Rules
4	Phone	Presidio + Rules
5	Fax	Rules (labeled)
6	Email	Presidio + Rules
7	SSN	Presidio + Rules
8	MRN	Transformer + Rules
9	Health Plan ID	Rules (MBI, Member ID patterns)
10	Account Numbers	Presidio + Rules
11	License/DEA	Rules (DEA, Medical License)
12	Vehicle ID (VIN)	Rules
13	Device ID (UDI)	Rules
14	URLs	Presidio + Rules
15	IP Addresses	Presidio + Rules
16-18	Biometric/Photos/Other	N/A (non-text)

Dictionary-Based Detection

Dictionary	Entity Type	Source
Drug Names	DRUG	FDA NDC (~30K)
Hospital Names	HOSPITAL	CMS Provider (~6K)
Insurance Payers	HEALTH_PLAN	Bundled (~100)
Lab Tests	LAB_TEST	Bundled (~300)
Provider NPIs	NPI	NPPES (~8M, SQLite)

Quick Start

# Install
cd privplay
pip install -e ".[presidio]"
python -m spacy download en_core_web_sm

# Download dictionaries (required)
phi-train download all

# Generate synthetic training data
phi-train faker -n 50

# Check detection stack status
phi-train stack

# Scan for PHI/PII
phi-train scan

# Interactive review session
phi-train review

# Check progress
phi-train stats

# Run F1 evaluation
phi-train test

Commands

Command	Description
phi-train faker	Generate synthetic clinical documents
phi-train import <path>	Import documents from files
phi-train scan	Scan documents for PHI/PII
phi-train review	Interactive review session
phi-train stats	Show training progress
phi-train test	Run F1 evaluation
phi-train detect <text>	Detect PHI in single text
phi-train download <target>	Download dictionaries/NPI
phi-train stack	Show detection stack status
phi-train export <file>	Export corrections
phi-train reset	Reset database
phi-train config-show	Show configuration

Workflow

1. Install: pip install -e ".[presidio]" + spacy model
2. Download: phi-train download all (dictionaries + NPI)
3. Generate: phi-train faker -n 100
4. Scan: phi-train scan (full detection stack)
5. Review: phi-train review (approve/reject below 95%)
6. Evaluate: phi-train test (see F1 score)
7. Iterate: Keep reviewing to improve accuracy

Configuration

Config file: ~/.privplay/config.yaml

confidence_threshold: 0.95
context_window: 50

presidio:
  enabled: true
  score_threshold: 0.5

verification:
  provider: ollama
  ollama:
    url: http://localhost:11434
    model: phi3:mini

Entity Types

Identity: NAME_PERSON, NAME_PATIENT, NAME_PROVIDER, SSN, MRN, NPI, DEA_NUMBER, MEDICAL_LICENSE Contact: EMAIL, PHONE, FAX, ADDRESS, ZIP Temporal: DATE, DATE_DOB, AGE Financial: CREDIT_CARD, BANK_ACCOUNT Digital: IP_ADDRESS, URL, USERNAME, VIN, UDI Clinical: DRUG, LAB_TEST, HOSPITAL, HEALTH_PLAN, HEALTH_PLAN_ID