Presidio Anonymizer

Warnings

• breaking The default behavior of the 'hash' operator changed in version 2.2.361. It now uses a random salt by default for enhanced security, which means the same PII value will yield different hashes across calls or entities. This breaks referential integrity unless a salt is explicitly provided.
  Fix: If referential integrity (same hash for the same value) is required, you must explicitly provide a consistent 'salt' parameter to the 'hash' operator config. Use a secure method to generate and store this salt.
• gotcha When using `presidio-analyzer` and `presidio-anonymizer` together in a typed Python environment, `mypy` might report type errors due to `RecognizerResult` existing in both packages with incompatible types. The `AnonymizerEngine` expects `RecognizerResult` from `presidio_anonymizer.entities`.
  Fix: Always import `RecognizerResult` specifically from `presidio_anonymizer.entities` when passing results to `AnonymizerEngine`. You may need to cast or convert `presidio-analyzer`'s `RecognizerResult` objects if you generate them from `AnalyzerEngine` and need to satisfy strict type checking.
• gotcha Starting from version 2.2.359, many country-specific recognizers (e.g., SgFinRecognizer, AuAbnRecognizer) are disabled by default to prevent false positives when they are not explicitly needed. Users expecting these to work out-of-the-box might find them inactive.
  Fix: If specific country-specific recognizers are required, ensure they are explicitly enabled or configured within your `presidio-analyzer` setup.
• gotcha In versions prior to 2.2.362, `AnonymizerEngine` could fail to anonymize all instances of an entity if multiple identical entities were separated only by spaces, potentially leading to PII leakage (e.g., 'email1@example.com email2@example.com'). A fix was implemented in 2.2.362.
  Fix: Upgrade to version 2.2.362 or later. Ensure thorough testing of anonymization output, especially for texts with repeated PII entities separated by whitespace.
• gotcha While `presidio-anonymizer` is a standalone package, a complete PII detection and anonymization pipeline typically requires `presidio-analyzer` for detection and an underlying NLP engine (like spaCy with a language model such as `en_core_web_lg`). Not installing these dependencies will prevent the full workflow from functioning.
  Fix: For a complete solution, install `presidio-analyzer` and a spaCy language model using `pip install presidio-analyzer "spacy[en]"` and `python -m spacy download en_core_web_lg`.

Install

• pip install presidio-anonymizer Install core anonymizer
• pip install presidio-analyzer "spacy[en]" python -m spacy download en_core_web_lg Recommended for full PII detection & anonymization

Imports

• AnonymizerEngine

  from presidio_anonymizer import AnonymizerEngine

• RecognizerResult

  from presidio_anonymizer.entities import RecognizerResult

  While presidio-analyzer also has a RecognizerResult, it's a separate type. For direct use with AnonymizerEngine, import from `presidio_anonymizer.entities` to avoid mypy type errors.
• OperatorConfig

  from presidio_anonymizer.entities import OperatorConfig

Quickstart

This quickstart demonstrates how to use `AnonymizerEngine` to anonymize text. It manually defines `RecognizerResult` objects, which would typically be generated by `presidio-analyzer`. It shows custom operators for different entity types.

from presidio_anonymizer import AnonymizerEngine
from presidio_anonymizer.entities import RecognizerResult, OperatorConfig

# Sample text and mock analyzer results (typically from presidio-analyzer)
text = "My name is John Doe and my phone number is 123-456-7890."
analyzer_results = [
    RecognizerResult(entity_type="PERSON", start=11, end=19, score=0.9),
    RecognizerResult(entity_type="PHONE_NUMBER", start=38, end=50, score=0.8),
]

# Initialize the anonymizer engine
anonymizer = AnonymizerEngine()

# Define anonymization operators
# Here, PERSON will be replaced with "<PERSON>", and PHONE_NUMBER will be masked
operators = {
    "PERSON": OperatorConfig("replace", {"new_value": "<PERSON>"}),
    "PHONE_NUMBER": OperatorConfig("mask", {
        "masking_char": "*", 
        "chars_to_mask": 10, 
        "from_end": True
    })
}

# Perform anonymization
anonymized_result = anonymizer.anonymize(
    text=text,
    analyzer_results=analyzer_results,
    operators=operators
)

print(f"Original text: {text}")
print(f"Anonymized text: {anonymized_result.text}")