Title: LLHR: Low Latency and High Reliability CNN Distributed Inference for Resource-Constrained UAV Swarms

Abstract: Recently, Unmanned Aerial Vehicles (UAVs) have shown impressive performance in many critical applications, such as surveillance, search and rescue operations, environmental monitoring, etc. In many of these applications, the UAVs capture images as well as other sensory data and then send the data processing requests to remote servers. Nevertheless, this approach is not always practical in real-time-based applications due to unstable connections, limited bandwidth, limited energy, and strict end-to-end latency. One promising solution is to divide the inference requests into subtasks that can be distributed among UAVs in a swarm based on the available resources. Moreover, these tasks create intermediate results that need to be transmitted reliably as the swarm moves to cover the area. Our system model deals with real-time requests, aiming to find the optimal transmission power that guarantees higher reliability and low latency. We formulate the Low Latency and High-Reliability (LLHR) distributed inference as an optimization problem, and due to the complexity of the problem, we divide it into three subproblems. In the first subproblem, we find the optimal transmit power of the connected UAVs with guaranteed transmission reliability. The second subproblem aims to find the optimal positions of the UAVs in the grid, while the last subproblem finds the optimal placement of the CNN layers in the available UAVs. We conduct extensive simulations and compare our work to two baseline models demonstrating that our model outperforms the competing models.