Researchers build an encrypted routing layer for private AI inference

This report summarizes SecureRouter, a University of Central Florida research system that enables private, input-adaptive routing for encrypted AI inference using secure multi-party computation (MPC). SecureRouter evaluates encrypted queries to choose an appropriately sized model from a pool—keeping routing decisions encrypted—to reduce inference latency while preserving client privacy; tests show average speedups of ~1.95× versus a fixed large-model private inference baseline with minimal accuracy loss and modest routing overhead.