Security-Oriented UAV Deployment, DNN Splitting, and Resource Allocation for Collaborative Inference

초록

Unmanned aerial vehicles (UAVs) as edge servers hold the potential to provide intelligent inference services using deep neural networks (DNNs). This letter focuses on a UAV-assisted end-edge collaborative inference system, where a DNN is split between each ground device and a UAV-mounted edge server for achieving co-inference. To secure the transmission of intermediate features extracted from the initial layers of the DNN, we propose a cooperative jamming scheme to disrupt eavesdropping. Our goal is to minimize the total inference delay by optimizing the UAV's deployment, DNN splitting, devices' transmit power, jamming power, and computation resource allocation. Since the problem is a mixed-integer nonlinear programming problem, we propose an efficient iterative algorithm that decomposes the problem into three subproblems. Specifically, we adopt a successive convex approximation method for UAV deployment and resource allocation subproblems, while a particle swarm optimization algorithm addresses DNN splitting. The subproblems are then alternatively optimized by a block coordinate descent approach. Numerical results demonstrate that the proposed scheme outperforms baselines in reducing the inference delay.

키워드