Airbnb Uses Presidio and GLiNER to De-identify PII in AI Training Data
·2026.11.18 03:01
Key point
Airbnb built a shared de-identification pipeline using Presidio and GLiNER to replace personal data with realistic, run-consistent fakes before ML training.
Details
- Airbnb implements a text de-identification system to remove PII from messages, reviews, and support chats before using them to train machine learning models.
- The system is built on the open-source Presidio library, hosted on-premises as a Python package to keep data on Airbnb's infrastructure and allow independent job execution.
- Detection uses a mix of pattern-based recognizers for IDs and NER-based recognizers; Airbnb replaced the default spaCy engine with GLiNER for PERSON and LOCATION entities to improve precision on specific data types like street addresses.
- The Anonymizer replaces detected PII with realistic fake values using the faker library, ensuring the same real value maps to the same fake within a single run but different fakes across separate runs to prevent fingerprinting.
- Quality is evaluated on recall, precision, and throughput, with metrics reported per entity type and language to ensure no regression in detection capabilities before releasing new versions.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
AAirbnb Introduces Proximity Features for Cold-Start Personalization Without User Identity
Spring AI Privacy Guardrails adds PII tokenization and tool authorizationAProject Lighthouse Part 3: Introducing project-lighthouse-anonymize
Toss TASTE Cuts Model Training to One Day and Automates Comparative Evaluation
Research finds 24,664 Lovable apps exposing sensitive data without login
UMAP 0.6 Release: 4x Faster Neighbour Search and Recursive Initialization for Large-Scale Embeddings