The evolution of sixth-generation (6G) networks enables transformative edge Artificial Intelligence of Things (AIoT) applications but introduces critical security vulnerabilities during model transmission between the central server and edge devices (e.g., unmanned aerial vehicles). Traditional approaches fail to jointly optimize model accuracy and physical layer security against eavesdropping. To address this gap, we propose a novel dynamic user selection framework that integrates three key innovations: (1) closed-form secrecy outage probability derivation for Rayleigh fading channels, (2) a Secure Model Accuracy (SMA) metric unifying recognition accuracy and secrecy outage probability, and (3) an alternating optimization algorithm for joint model–bandwidth selection under secrecy constraints. Comprehensive simulations demonstrate 22% SMA gains over baselines across diverse channel conditions and eavesdropper capabilities, resolving the fundamental accuracy–security tradeoff for trustworthy edge intelligence.