Deep neural network inference task offloading and resource allocation for UAV-based wireless networks.

Ismail, Muhammad (2025). Deep neural network inference task offloading and resource allocation for UAV-based wireless networks. Mémoire. Québec, Université du Québec, Institut national de la recherche scientifique, Maitrise en télécommunications, 103 p.

Prévisualisation

PDF - Version publiée Télécharger (1MB) | Prévisualisation

Résumé

Novel integration of Deep Neural Networks (DNNs) into Unmanned Aerial Vehicles (UAVs) based systems has enabled numerous smart civilian and military applications. However, the inherent computational complexity of DNNs often results in significant inference delay, which could violate the delay requirements of practical UAV operations and applications. UAVs can be equipped with relatively strong servers so they can collaboratively perform inference for pre-trained DNNs, en-abling complex recognition tasks based on onboard sensing data such as image and video. Such the collaborative inference is critical for applications where ground communications and comput-ing infrastructure is not available, not secure or cost-efficient such as those for military, disaster recovery and rescue. Collaborative DNN inference in the UAV wireless network, is, however, chal-lenging because one must decide how the computation load related to different layers of the DNN is distributed among UAVs and how to efficiently allocate both radio and computing resources to facilitate the underlying offloading process. This thesis aims to address these challenges where we make the following novel contributions.

First, we formulate the joint DNN layer assignment, radio and computing resource allocation problem as an optimization problem which aims to minimize the total inference latency considering constraints on the processing order of DNN layers, wireless connectivity, and limited computational resources of individual UAVs. To solve this difficult mixed integer and non-linear problem, we employ an alternating optimization technique and develop an efficient algorithm, named LARA. Numerical studies show that LARA performs very well in different studied scenarios and achieves up to 80% improvement in terms of inference latency compared to other baselines which perform DNN layer assignment and resource allocation in a heuristic manner.

Second, we propose a novel decentralized matching game theory-based algorithm (called GAME-MIND), enabling efficient allocation of DNN layers of different inference requests to the UAVs considering UAVs’ processing quotas. By modeling the problem as a matching game, GAME-MIND ensures efficient resource utilization while balancing the computation load across UAVs in the network. Numerical results show that GAME-MIND achieves desirable delay performance compared to the centralized LARA counterpart. In addition, GAME-MIND greatly outperforms two different heuristic-based methods that offload DNN layers to nearest neighboring UAVs where these two heuristic baselines can achieve at least twice the inference delay due to GAME-MIND in high-load conditions. We also demonstrate the efficacy of GAME-MIND in terms of convergence and study its achieved communications, computation latency, and total inference delay under different network settings.

Type de document:	Thèse Mémoire
Directeur de mémoire/thèse:	Le, Long Bao
Mots-clés libres:	-
Centre:	Centre Énergie Matériaux Télécommunications
Date de dépôt:	01 mai 2026 18:43
Dernière modification:	01 mai 2026 18:43
URI:	https://espace.inrs.ca/id/eprint/17138

Gestion Actions (Identification requise)

Modifier la notice