Harnessing LLMs for Scalable Data Minimization
summary
Claims that foundation LLMs outperform regex for messy column formats by labeling table/column/examples into categories like name phone email other, and that coding agents can trace a savePhoneNumber call through files to a DB cache and logged event and draft diffs to remove unwanted flows. Experiments show higher contextual PII recall, useful at org scale, but higher latency and cost, risks sending examples externally, context length caps, and hallucination false positives. Proposed pipeline stitches a PII detector, a privacy coding agent, and an org context assistant into a retention watchdog.
AI-written summary · about
chapters (6)
- 0:00Data minimization with LLMs
- 1:59LLMs classify PII
- 6:10LLMs map data flows
- 11:07LLMs as context assistants
- 12:54Automated retention watchdog
- 14:17Summary and future
related talks
33mDesigning LLM Applications for Massive Datasets
56mDehumanizing Agents: Why Explainability is Crucial in the LLM Era
29mAutomating Privacy Reviews with LLMs: Challenges & Opportunities
28mDon't leak user data to AI - Strategies for protecting PII from LLMs and MCP
22mPractical Strategies for Implementing Domain-Specific LLMs
31mLeveraging LLMs to build supervised datasets suitable for smaller models
06mData Privacy Concerns and Solutions When using LLMs in Production
36mTechniques to avoid common LLM pitfalls