Balancing Privacy & Utility: Efficient PII Detection and Replacement in Textual Data
summary
Pipeline claims a practical, open‑source two‑step solution: detect PII with pre‑trained NER (gliner) then replace with controlled synthetic data via Faker, preserving layout and statistical traits. Fine‑tuning uses 800 LLM‑assisted fake résumés; results show improved entity scores except first/last name split, where switching to a person label plus a binary first/last classifier helps. Replacements enforce length, casing, gender, and do in‑place PDF edits; no real candidate data is sent to LLM providers.
AI-written summary · about
chapters (9)
- 0:00PII detection basics
- 4:41Two-step anonymization pipeline
- 7:09Evaluating PII models
- 10:42Fine-tuning with resume data
- 15:31Improving name detection
- 17:39Replacing PII realistically
- 20:10Customizing fake data
- 24:52Key takeaways
- 27:01Q&A
related talks
01h 27mLLM Privacy Paradox: Balancing Data Utility...
20mPreserving Personal Data Privacy in the Age of AI: A Practical Approach with Nerpii
28mAnonymization in the LLM Era: No Room for Compromise
23mSynthetic data that meets enterprise compliance needs
29mPersonal Data Pseudonymization Versus Anonymization In The Age Of AI & Big Data
03mIT Security Summit 2025 | Pablo Llano | Nymiz
28mInsights, Not Secrets: How Synthetic Data Empowers Privacy
27mWhat’s in a Name? – Fuzzy Matching Techniques for Proper Nouns