Joint Split Inference Design and Bandwidth Allocation for Intelligent Edge Perception Network

The integration of artificial intelligence and wireless sensing at the network edge offers significant potential to alleviate device-side computational burdens and reduce the computational delay of the sensing task. However, offloading raw sensing data to Mobile Edge Computing (MEC) servers often leads to substantial transmission delays. To address this, we introduce split inference into edge perception networks and propose a three-tier collaborative inference architecture involving cloud servers, MEC servers, and devices. Each tier executes a portion of a Deep Neural Network (DNN)-based sensing task. Leveraging Integrated Sensing and Communication (ISAC) technology, devices simultaneously sense the environment and offload intermediate DNN features to the MEC or cloud server for further inference. We formulate a joint optimization problem to determine the adaptive DNN splitting strategy and bandwidth allocation, aiming to minimize total inference delay. To solve the resulting mixed-integer nonlinear programming problem, we develop a two-layer optimization algorithm based on the Karush-Kuhn-Tucker conditions and the cross-entropy method. Simulation results show that the proposed framework significantly reduces average inference delay compared to existing two-tier collaborative inference schemes.