Dynamic DNN Model Switching for Collaborative Edge Intelligence in UAV Swarm

Deploying machine learning (ML) algorithms on Unmanned Aerial Vehicle (UAV) swarm to perform real-time tasks such as object detection is currently a hot research topic. However, as the parameter size of deep neural network (DNN) models continues to grow rapidly, deploying complex DNN models on resource-constrained UAV swarm poses significant challenges. In this paper, we first study how to dynamically switch between small lightweight DNN model and large complex DNN model for efficient edge intelligence in a UAV swarm, where the former is executed by a single UAV and the latter executed by multiple UAVs collaboratively. Specifically, we formulate an optimization problem to maximize the utility capturing the tradeoff between the average accuracy and latency per DNN inference request, with the constraints of computation and memory, under the time-varying UAV-to-UAV communication links. To address it, we propose an algorithm called Dynamic Model Switching (DMS) based on genetic algorithms to select the best DNN model switching stragety, according to the current status of UAV communication network. Trace-driven simulation results validate the effectiveness of the proposed DMS algorithm by always achieving the highest utility value when varying number of DNN inference requests or number of UAVs, compared to four baseline algorithms.