Anonymization in the LLM Era: No Room for Compromise
summary
LLMs amplify data leaks: an employee once uploads 3,000 records uploaded and an LLM-powered pipeline multiplies exposure across microservices, RAG, agents and autonomous tool calls. Traditional NER is 2,000x slower vs machine learning for throughput?—no, LLMs are orders of magnitude slower and nondeterministic; reasoning often hurts extraction. Replace PII with structured placeholders tied to a secure token store so you can rehydrate or mask per-access rules; naive redaction fails — even a title leak reidentified original after all entities were replaced.
AI-written summary · about
chapters (13)
- 0:00Introduction and speaker intro
- 1:25Privacy challenges with LLMs
- 3:37Evolution of LLM ecosystem complexity
- 5:10Demo setup and document example
- 6:14Traditional PII extraction methods
- 8:22LLM-based entity extraction drawbacks
- 10:59Replacing entities with placeholders
- 13:51Risks of LLM-generated data in production
- 15:39Entity resolution and access control
- 19:57Handling dates, times, and addresses
- 23:29Why anonymization is hard
- 24:59Hidden risks in document titles
- 26:25Conclusion and no room for compromise
related talks
28mDon't leak user data to AI - Strategies for protecting PII from LLMs and MCP
06mData Privacy Concerns and Solutions When using LLMs in Production
44mI Abuse Your LLM Live with 10 Security Vulnerabilities
01h 27mLLM Privacy Paradox: Balancing Data Utility...
20mSecurity Threats in Modern LLM Applications
23mSynthetic data that meets enterprise compliance needs
29mAutomating Privacy Reviews with LLMs: Challenges & Opportunities
08mLLMs Without Leaks: Keep Your Code, Data & IP