Polski
Open Source · Apache-2.0 v1.1

PII Masker

Multi-tier PII anonymization and privacy protection engine for AI pipelines

Online Tester (masker.apps) GitHub: anonymizer-model GitHub: fast_masker plugin
  • PII Masking
  • Data Privacy
  • NER Model
  • FastMasker
  • GDPR / Privacy
  • LLM Security

Why PII Protection is Critical in AI Workflows

Sending user queries, enterprise documents, and customer tickets to public cloud LLMs creates significant risks of data leakage, confidentiality breaches, and GDPR/compliance violations.

Traditional regex solutions struggle with high false-positive rates or miss non-standard formatting, while pure deep-learning models often fail to strictly validate digital identifiers that follow formal checksum algorithms.

PII Masker addresses both challenges through a hybrid two-tier architecture: it couples high-certainty deterministic checksum rules with specialized Named Entity Recognition (NER) models for unstructured Polish and multilingual text.

Two-Tier Hybrid Architecture

Raw text with sensitive information

  1. 1. FastMasker

    Deterministic rules and checksum validation: PESEL, NIP, REGON, IBAN, credit cards and VIN. Format rules: email, IPv4/IPv6, URLs, phone numbers and postcodes.

  2. 2. Anonymizer Model

    Transformer NER token classification: personal names, organizations and roles, as well as addresses, locations and contextual entities.

Anonymized text with structured placeholders, e.g. {{PESEL}}, {{CREDIT_CARD}}

Core Components

1. FastMasker: Ultra-Fast Rule Engine

The fast_masker plugin applies validation rules in strict precedence order — from highest certainty (checksum-verified identifiers) down to generalized patterns:

  • Checksum-Validated Identifiers:
  • Credit Cards (CreditCardRule): Luhn algorithm checksum verification (13–19 digits) → {{CREDIT_CARD}}
  • Vehicle Identification Numbers (VinRule): ISO 3779 checksum (position 9) → {{VIN}}
  • Polish PESEL (PeselTaggedRule, PeselRule): Formal checksum calculation and birthdate sanity check → {{PESEL}}
  • Polish Tax (NIP) & Business (REGON) IDs: Weighted checksum validators → {{NIP}}, {{REGON}}
  • Bank Accounts (IbanRule, NrbRule): Polish NRB (26 digits) and international IBAN (modulo 97) → {{IBAN}}
  • Identity Documents: Polish National ID (IdCardNumberRule) and Passport numbers (PassportNumberRule) → {{ID_CARD}}, {{PASSPORT}}
  • System Credentials: MAC addresses, SIM ICCIDs, SSL serials, and JWT tokens → {{MAC_ADDRESS}}, {{JWT}}
  • Pattern-Based Identifiers:
  • Email addresses → {{EMAIL}}
  • IPv4 and IPv6 addresses → {{IP_ADDRESS}}
  • URLs and domains → {{URL}}
  • Polish and international telephone numbers → {{PHONE}}
  • Postal codes, license plates, monetary amounts, and dates

2. Anonymizer Model: ML NER for Polish Text

For unstructured entities without deterministic checksums (first and last names, street names, organization roles), we provide the anonymizer-model:

  • Architecture: RoBERTa-based AutoModelForTokenClassification tuned specifically for Polish syntax.
  • Trained on Curated Datasets: Built on annotated corpora such as clarin-pl/kpwr-ner with generalized entity taxonomies.
  • Advanced Post-Processing: Sub-token boundary stitching, punctuation preservation, and exact entity boundary alignment.
  • Low-Resource Deployment: Dynamic quantization (ONNX / INT8) enables fast CPU inference without requiring dedicated GPUs.

LLM Router Integration

PII Masker integrates seamlessly into LLM Router:

  1. Pre-Inference: Replaces all confidential identifiers in the prompt with structured placeholders.
  2. Inference: Cloud or local models generate responses based on sanitized data.
  3. Post-Inference (Optional): Restores original entities into the response before delivering text back to the authorized user.

Python Quickstart 

from llm_router_plugins.maskers.fast_masker import FastMasker, ALL_RULES

# Initialize the masker with all built-in rules
masker = FastMasker(rules=ALL_RULES)

text = (
    "Customer John Doe, PESEL: 44051401359, paid the invoice using card "
    "4532 1234 5678 9010 for 1,250.00 USD. "
    "Contact: john.doe@company.com or +48 601 234 567."
)

masked_text = masker.mask(text)
print(masked_text)
# Output:
# Customer John Doe, {{PESEL_TAGGED}}, paid the invoice using card
# {{CREDIT_CARD}} for {{MONEY}}.
# Contact: {{EMAIL}} or {{PHONE}}.

Resources & Repositories