Presidio Anonymizer

Presidio Anonymizer is the de-identification half of the Presidio data-protection toolkit, taking the spans of personally identifiable information found by an upstream detector (typically Presidio Analyzer) and applying an operator — replace, redact, mask, hash, encrypt, or a custom lambda — to each one. It also ships Deanonymizers that reverse a prior anonymization, such as decrypting text that was encrypted with a known AES key, which lets teams round-trip sensitive data through a pipeline without ever storing the plaintext.

The package is deliberately narrow in scope: it does not detect PII itself, it only transforms already-identified entity spans. That separation keeps it usable as a standalone text-rewriting engine wherever a caller already has offsets and entity types, whether those come from Presidio Analyzer, a third-party NER model, or hand-rolled regex, while still composing cleanly with the rest of the Presidio stack for a full identify-then-anonymize workflow.

What You Get

• An AnonymizerEngine that applies per-entity-type operators (replace, redact, mask, hash, encrypt, custom lambda, keep) to a text plus a list of recognizer results
• A DeanonymizeEngine that reverses encryption-based anonymization given the original key, enabling round-trip workflows
• A BatchAnonymizerEngine / BatchDeanonymizeEngine for processing dict-shaped batches of text fields in one call
• An AES-based Encrypt/Decrypt operator pair plus Hash (SHA-256/SHA-512), Mask, Redact, and Replace operators out of the box
• An optional AHDS Surrogate operator that calls Azure Health Data Services de-identification to generate realistic, format-preserving surrogates for PHI
• Conflict-resolution logic for overlapping PII spans (highest score wins, configurable strategies for equal-score ties)

Common Use Cases

• Anonymizing chat transcripts, support tickets, or logs before sending them to an LLM or third-party analytics service
• Encrypting detected PII in place so it can be securely deanonymized later by an authorized process holding the key
• Masking or redacting sensitive fields in structured records ahead of data sharing or export
• Building a compliance pipeline where Presidio Analyzer finds entities and Presidio Anonymizer enforces the redaction policy per entity type

Under The Hood

Architecture - AnonymizerEngine.anonymize() takes raw text plus a list of RecognizerResult spans, resolves overlapping spans via ConflictResolutionStrategy, and hands each surviving span to the operator registered for its entity type through OperatorsFactory, using a TextReplaceBuilder to rewrite the string span-by-span without corrupting downstream offsets; DeanonymizeEngine mirrors this for reversing Encrypt operations. Tech Stack - pure Python with cryptography for AES encryption/decryption as its only hard dependency; optional extras add Flask/Gunicorn/Waitress for a standalone HTTP server and azure-identity/azure-health-deidentification for the AHDS surrogate operator; packaged with Poetry. Code Quality - a tests/ directory with roughly 32 test modules covering each operator, conflict resolution, and the batch engines individually, plus typed dataclass-style entities (InvalidParamException, RecognizerResult) for input validation with explicit error messages rather than silent failures. API Design - the operator model (OperatorConfig(operator_name, params) keyed per entity type) is consistent across all built-in and custom operators, keeping the public surface small: instantiate an engine, pass text plus analyzer results, get back anonymized text and applied-operator metadata.