The ChatGPT Privacy Tango: Dancing with Data Security and Large Language Models
summary
A working proxy system uses AWS Nitro enclaves and mutual attestation to encrypt uploads, run a de‑identification model inside enclaves, build embeddings into a vector DB, send redacted prompts to LLM providers, and re‑identify masked tokens with a saved mapping the demo returns sensible answers and shows the flow end to end the de‑identification model reports about 99% accuracy but misses coreference and reidentification attacks remain possible so privacy increases but is not formally differentially private
AI-written summary · about
chapters (15)
- 0:00Intro
- 1:00Cape Privacy
- 2:44Overview
- 3:51Large Language Models
- 8:38Privacy Concerns with LLMs
- 11:50ChatGPT Privacy Tango
- 15:38Confidential Computing
- 18:17PII Removal and Text De-identification
- 22:07What the client sees
- 30:02System Design: Setup
- 31:50System Design: Knowledge Base
- 32:58System Design: User Query
- 38:02System Analysis: Benefits
- 45:04System Analysis: Disadvantages
- 47:16Alternative approaches
related talks
14mToward Provably Private Insights into AI Use
25mSafeguarding Sensitive Data in the Age of Generative AI
28mPrivacy preserving machine learning
32mPrivate Insights into AI Use
39mData Privacy Engineering to power Responsible Enterprise AI
51mSecuring LLM-Powered Applications: Overcoming Security and Privacy Challenges
44mSecuring LLM-Powered Apps: Overcoming Security and Privacy Challenges
21mChatGPT, LLMs, and LangChains - A Beginner's Guide