Task-Oriented Multi-User Wireless Model Delivery for Edge-Assisted Autonomous Driving

Abstract

Edge-assisted autonomous driving increasingly depends on delivering task-specific deep neural network (DNN) models from the infrastructure to vehicles, so that the environment perception models can be updated to adapt to changing traffic environments, service requirements, and heterogeneous onboard platforms. However, transmission errors may degrade the performance of the supported task, which cannot be directly reflected by the conventional communication performance metrics such as bit error rate (BER). This paper studies a task-oriented wireless model delivery process for quantized DNNs in a multi-user scenario. Using validation-loss increment as a unified indicator for heterogeneous perception tasks, we establish a tractable BER-to-loss function by relating BER-induced quantized-parameter perturbations to validation-loss increment through parameter sensitivity. On this basis, a resource allocation strategy is developed under delay and energy constraints to improve the reliability of delivered models. Extensive simulation results show that the proposed solution achieves lower task degradation than the classic communication-centric baselines under tight system budgets. These results highlight the need to account for model sensitivity in wireless model delivery for edge-assisted autonomous driving tasks.