Resilient Priority-Aware Inference for Mission-Critical UAV Networks Under Bursty Edge Load

Abstract

Mission-critical mobile networks, including UAV-based disaster-response systems, must keep delivering usable and trustworthy inference under the exact conditions that make inference harder: constrained edge compute, degraded backhaul, and requests arriving in bursts rather than at a steady rate. A system that allocates compute uniformly has no way to fail gracefully once a burst exceeds its budget, dropping high-urgency requests along with low-urgency ones. In this paper, we present a priority-aware admission-control architecture built around two trust mechanisms: a lightweight router that assigns each request a priority level grounded in verifiable evidence rather than a model confidence, and a deterministic safety gate that rejects unsafe outputs regardless of which model checkpoint produced them. Low- and medium-priority requests are served by a baseline checkpoint, and a confidence-gated escalation path that reserves a more capable checkpoint for requests that need it. Both mechanisms are designed as auditable security controls rather than opaque model judgments: the router functions as a content-based access-control decision, and the safety gate as a policy-enforcement point that cannot be bypassed by manipulating the model confidence. We evaluate this policy end-to-end. The priority queue degrades gracefully, exactly as designed, under a simulated burst load. The router and confidence gate that feed it, however, correctly protect only 46.58% of genuinely high-priority requests, and a policy-optimization stage did not measurably improve the underlying checkpoint. Ultimately, a safety mechanism is only as trustworthy as the evidence behind it.