This page may contain stale information. Last updated: 2026-08-15

Definition

Inference pipelines that use privacy proxies (oblivious-http, masque, TEEs, etc.) so model hosts or relays cannot jointly observe user identity and query content — increasingly relevant for health/finance and agent-mediated access.

Key Points

HE/FHE inference typically carries 100x–1000x+ latency overhead vs plaintext — suited for regulated niches, not general LLM chat at scale.

Sources