Zero Trust Architectures for Edge-Native AI Inference Systems

Abstract

Edge‑Native AI Inference Systems (ENAIS) are rapidly proliferating across domains such as autonomous vehicles, smart manufacturing, healthcare diagnostics, and critical infrastructure monitoring. By bringing AI inference closer to data sources, ENAIS dramatically reduce latency, conserve bandwidth, and enable real‑time decision‑making. However, the shift from centralized cloud environments to widely distributed, resource‑constrained edge nodes introduces unique security challenges: traditional perimeter defenses become ineffective, attack surfaces multiply, and heterogeneity complicates policy enforcement. To address these concerns, this manuscript presents a comprehensive Zero Trust Architecture (ZTA) tailored for ENAIS. Building on the principle of “never trust, always verify,” our design integrates three core components: identity‑centric access controls, micro‑segmentation of inference pipelines, and continuous telemetry‑driven policy adaptation. We detail the architectural blueprint, describe its implementation within a simulation framework, and conduct a rigorous evaluation under realistic threat scenarios. Statistical analysis of the simulation data—covering metrics such as unauthorized access prevention, inference latency, and resource overhead—reveals that the proposed ZTA blocks over 98% of unauthorized actions while incurring only a modest 16.6% latency penalty and minimal CPU and network overhead. These results demonstrate that ZTA can substantially elevate the security posture of ENAIS without compromising real‑time performance requirements. We conclude with a discussion of deployment considerations, potential integration with federated learning, and directions for future work in securing next‑generation edge AI.