Low Latency and Accuracy-Guaranteed DNN Inference for UAV-Assisted IoT Networks
Xu, Jiaqi Yao, Haipeng Zhang, Ru Mai, Tianle Guizani, Mohsen
Xu, Jiaqi
Yao, Haipeng
Zhang, Ru
Mai, Tianle
Guizani, Mohsen
Supervisor
Department
Machine Learning
Embargo End Date
Type
Journal article
Date
2025
License
Language
English
Collections
Research Projects
Organizational Units
Journal Issue
Abstract
In recent years, unmanned aerial vehicles (UAVs) have shown great potential in various Internet of Things (IoT) applications, such as monitoring, inspection, and data collection. Especially, by integrating deep neural network (DNN) technology, UAV networks can provide real-time data insights, enhancing decision-making processes and ultimately boosting industrial productivity while saving costs. In this context, the environmental monitoring data collected by task nodes requires further inference, thereby generating a continuous stream of inference tasks. These tasks should be scheduled and offloaded to UAVs equipped with DNN models, i.e., inference nodes, for further real-time inference. However, the deployment of DNN models requires substantial computing resources, posing challenges for resource-constrained UAV networks. Fortunately, the knowledge distillation (KD) technique offers an effective solution by deriving lightweight models that maintain high accuracy. In this way, the UAV network can request and deploy the proper lightweight model for DNN inference. To attain efficient DNN inference, the inference task assignment, DNN model selection, and bandwidth allocation are coordinately considered to minimize the average inference delay while satisfying the strict accuracy requirements of inference tasks. In this paper, we formulate this problem as a Constrained Markov Decision Process (CMDP) and propose a Constrained Deep Reinforcement Learning based Task Allocation and Resource Management (CDRL-TR) algorithm to dynamically obtain solutions. Furthermore, a resource allocation optimization subroutine is embedded in the proposed algorithm to accelerate the training process. Extensive simulations demonstrate that the CDRL-TR algorithm has satisfactory performance in terms of task accuracy guarantee and inference delay optimization.
Citation
J. Xu, H. Yao, R. Zhang, T. Mai and M. Guizani, "Low Latency and Accuracy-Guaranteed DNN Inference for UAV-Assisted IoT Networks," in IEEE Transactions on Cognitive Communications and Networking, doi: 10.1109/TCCN.2025.3542443
Source
IEEE Transactions on Cognitive Communications and Networking
Conference
Keywords
Autonomous aerial vehicles, Artificial neural networks, Inference algorithms, Resource management, Optimization, Computational modeling, Heuristic algorithms, Real-time systems, Accuracy, Internet of Things
Subjects
Source
Publisher
IEEE