New research compares PII redaction methods for LLM pipelines

A new research paper compares four methods for redacting personally identifiable information (PII) before sending text to large language models (LLMs). The study evaluated pure regular expressions (regex), Microsoft Presidio, and two hybrid approaches. While regex was the fastest, it only achieved 88.9% PII redaction accuracy. The hybrid methods, including one that directly uses spaCy's named-entity component, achieved 100% PII redaction and preserved business terms, with the fastest hybrid adding only a minimal 0.41% latency to a typical LLM call. AI

IMPACT The fast hybrid redaction method is recommended for LLM pipelines handling personal data, balancing speed and accuracy.

RANK_REASON The cluster contains a research paper evaluating different technical methods for data redaction. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• LLMs
• Microsoft Presidio
• Pii
• Presidio
• ProfileAssist
• regular expression
• SpaCy

• paper
• safety

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →