Safeguarding Sensitive Data in the Age of Generative AI
summary
Claims that LLMs have no delete button and that sensitive data must be fenced with a privacy vault drive the work here. A medical customer‑support assistant is built: Next.js frontend, Django backend, OpenAI foundation model augmented with ChromaDB embeddings via LangChain, ETL into Snowflake, and a Skylow proxy that tokenizes or deidentified data for training and inference and then reidentifies outputs under fine‑grain access control. Live UI clicks fail but the pipeline logic and policy enforcement prove the approach viable.
AI-written summary · about
chapters (9)
- 0:00Introduction to talk
- 1:28Basics of large language models
- 3:34Historical tech paradigm shifts
- 8:53LLM privacy challenges
- 12:47Data privacy Vault concept
- 15:39Privacy in model training
- 17:04Privacy in model inference
- 18:37Medical support demo overview
- 23:52Key takeaways and conclusion
related talks
45mCombating Generative AI's Privacy Abuses
19mKeeping people away from data: A generative AI use case
40mSecuring Generative AI: Is it all an Illusion?
09mSecuring Generative AI Workloads
31mLeverage AI to Protect Your Data Assets... from AI
13mDeploying Enterprise-Grade Compliant Generative AI
43mSecurity of GenAI & LLMs: A New Era of Ethical Hacking
45mSecuring and Governing Generative AI: Learnings from Microsoft