Title: Adaptive offloading in multi-access edge networks via hierarchical federated learning and real-time system adaptation

Authors: Jie Wang; Qiao Liang; Amin Mohajer

Addresses: Big Data & Intelligence Engineering School, Chongqing College of International Business and Economics, Hechuan, Chongqing, 401520, China ' School of Mathematics and Computer Science, Chongqing College of International Business and Economics, Hechuan, Chongqing, 401520, China ' Department of Electrical and Computer Engineering, The University of British Columbia (UBC), Vancouver, BC V6T 1Z4, Canada

Abstract: Achieving ultra-reliable real-time digital twin (DT) adaptation in mobile edge environments requires intelligent orchestration of computation and communication under user heterogeneity and dynamic mobility. This paper introduces GADENet, a graph attention-enhanced digital twin evolution network that fuses graph neural modelling, multi-agent actor-critic learning, and hierarchical federated personalisation to enable seamless digital representations of user equipment (UE) in distributed edge networks. At its core, GADENet employs a GAT-assisted multi-agent deep deterministic policy gradient (MADDPG) framework to jointly learn optimal DT migration and personalisation strategies across edge servers, guided by real-time traffic topologies and resource interdependencies. Each DT model is modularised into generalisable and adaptive subspaces, trained collaboratively through a three-tier edge-cloud federated loop and refined using localised attention-based updates. For efficient mobility handling, we propose a parameter-sliced DT relay protocol that selectively migrates the minimal personalisation subset across servers, leveraging learned action-value functions to minimise response latency. Extensive simulations on CIFAR-based datasets and synthetic edge workloads demonstrate that GADENet achieves up to 30% reduction in interaction latency and significantly boosts modelling fidelity versus strong federated and DRL-based baselines. This work offers a principled blueprint for intelligent DT deployment under the constraints of 6G and next-gen IoT fabrics.

Keywords: digital twin modelling; graph attention networks; GATs; multi-agent deep reinforcement learning; federated learning; intelligent network orchestration.

DOI: 10.1504/IJSNET.2025.148455

International Journal of Sensor Networks, 2025 Vol.49 No.1, pp.1 - 17

Received: 28 Sep 2024
Accepted: 11 May 2025
Published online: 05 Sep 2025 *

Full-text access for editors Full-text access for subscribers Purchase this article Comment on this article