ai4privacy python module 🛡️

A Python package for state-of-the-art PII detection and masking using advanced transformer models.

Features

• Protect Mode: Anonymize text by replacing detected PII with placeholders.
• Observe Mode: Get statistics and a detailed "privacy mask" of found PII without altering the original text.
• Multiple Models: Built-in support for:
  • English-specific detection.
  • Multilingual detection.
  • Categorical detection (e.g., GIVENNAME, EMAIL, CITY).
• Tunable Sensitivity: An adjustable score_threshold to balance detection accuracy with false positives.
• Verbose & Developer Modes: Rich outputs for detailed analysis and debugging.

Installation

pip install ai4privacy

Quick Start

The simplest way to use the library is to call the protect function, which masks PII with placeholders.

from ai4privacy import protect

text = "Email me at developers@ai4privacy.com or call me at +41763223001."
masked_text = protect(text)

print(masked_text)
# Expected Output: Email me at [PII_1] or call me at [PII_2]

Advanced Usage

Using Different Models

You can easily switch between models using the multilingual and classify_pii flags.

from ai4privacy import protect

text = "Je m'appelle Pierre et j'habite à Paris."

# Use the multilingual model for non-English text
masked_multilingual = protect(text, multilingual=True)
print(f"Multilingual: {masked_multilingual}")
# Expected Output: Multilingual: Je m'appelle [PII_1] et j'habite à [PII_2]

# Use the categorical model to see the PII types
details = protect(text, classify_pii=True, verbose=True)
print(f"Categorical Labels: {[r['label'] for r in details['replacements']]}")
# Expected Output: Categorical Labels: ['GIVENNAME', 'CITY']

Observe Mode

To analyze text without changing it, use observe(). It returns a dictionary containing statistics and the privacy_mask—a detailed list of all PII entities found.

from ai4privacy import observe
import json

text = "My name is Alice and I live in Berlin."
report = observe(text, classify_pii=True)

print(json.dumps(report, indent=2))

{
  "num_texts_processed": 1,
  "num_texts_with_pii": 1,
  "pii_entity_counts": {
    "GIVENNAME": 1,
    "CITY": 1
  },
  "total_pii_entities_found": 2,
  "privacy_mask": [
    {
      "label": "GIVENNAME",
      "start": 11,
      "end": 16,
      "activation": 0.98,
      "value": "Alice"
    },
    {
      "label": "CITY",
      "start": 30,
      "end": 36,
      "activation": 0.99,
      "value": "Berlin"
    }
  ]
}

Verbose and Developer Modes

Set verbose=True to get a dictionary containing the original text, masked text, and replacement details. For deep debugging, developer_verbose=True adds a token-by-token breakdown of the model's predictions.

from ai4privacy import protect

text = "Senden Sie es an Eva Schmidt."
details = protect(text, classify_pii=True, verbose=True)

print(details['replacements'])
# Expected Output: [{'label': 'GIVENNAME', 'start': 18, 'end': 22, ...}, {'label': 'SURNAME', 'start': 23, 'end': 30, ...}]

Adjusting Sensitivity

The score_threshold (default: 0.01) controls how confident the model must be to flag a token as PII.

• A lower value increases sensitivity (finds more PII, but may have more false positives).
• A higher value increases precision (detections are more likely correct, but may miss some PII).

from ai4privacy import protect

text = "Maybe this is a name, maybe not. Contact John."

# High precision (less likely to flag "Maybe")
masked_high_prec = protect(text, score_threshold=0.5) 
print(f"High Precision: {masked_high_prec}")
# Expected Output: High Precision: Maybe this is a name, maybe not. Contact [PII_1]

# High sensitivity (more likely to flag "Maybe" if the model is unsure)
masked_high_sens = protect(text, score_threshold=0.001)
print(f"High Sensitivity: {masked_high_sens}")

Disclaimer 📢

Ai4Privacy is trained on the world's largest open-source privacy dataset. For production use, please evaluate results carefully on your own datasets. For assistance, contact us at our website https://ai4privacy.com or email support@ai4privacy.com.