Microsoft patents deletable user proxies for privacy-preserving LLM personalization
SchoeneggerPhil · x · 2026-11-10
Philipp Schoenegger received a Microsoft patent award cube for "Separable Expert Architecture," a paper on privacy-preserving LLM personalization with Chris Schneider and Ben Bariach.
Key idea: current training bakes user data into shared weights, making per-user deletion computationally infeasible. The three-layer architecture decouples personal data from shared weights via:
- a static base model plus composable domain-expert LoRA adapters that shape behavior without carrying user data
- per-user proxy artefacts whose deletion constitutes deterministic unlearning
Evaluated on Phi-3.5-mini and Llama-3.1-8B: outputs return to baseline after proxy removal (KL divergence 0.21 nats, 82–89% verification pass rate), with near-zero cross-user contamination. Since user-specific info never enters shared weights, the design mitigates model inversion, membership inference, and training-data extraction by construction, and is compatible with DP-SGD.
More from Safety
- USB+ open-sources a trust layer for inspecting agent skills and MCP servers before install — peepsick · 2026-11-10
- MCP server security hinged on $HOME — and the agent controls the environment — Relevant_Mousse_9117 · 2026-11-10
- Fireworks AI got hacked despite SOC 2 and ISO certs — 'security theater' debate — DesireeCachette · 2026-11-10
- Android's Accessibility permission can let apps read and interact with your screen — Aiden_Tech_Ai · 2026-11-10
- UK man arrested after Facebook flags AI-generated photo he sent wife on Messenger — neurocy · 2026-11-10
- Retro-Terminal 'Dream Man' Slides Steal the Show at WIPO's IP and AI Forum — technollama · 2026-11-10