Abstract
In next-generation intelligent networks, heterogeneous devices will operate in dynamic and untrusted environments. Machine learning-based task scheduling becomes increasingly susceptible to adversarial attacks, shifting trust landscapes. An urgent need arised for robust and trustworthy resource management mechanisms. To enable time-sensitive and secure task scheduling, we propose a Heterogeneous Graph Neural Network–augmented Deep Reinforcement Learning (HGNN-DRL) framework. The framework models multi-step, multi-device scheduling as a knowledge graph to account for various practical constraints. Based on the knowledge graph, an HGNN is designed to extract topological embeddings, which jointly capture task dependencies and device security states. To drive decision-making, we introduce an Entropy-Regularized Double Deep Q-Network (ER-DDQN) strategy, which decouples action selection from value estimation through the DDQN architecture. This decoupling improves training stability by separating target action selection from Q-value evaluation. Meanwhile, policy entropy is maximized to enhance exploration and improve robustness in dynamic environments. A Lyapunov-stability analysis shows that ER-DDQN ensures stable scheduling performance under dual constraints of delay and security. Simulations demonstrate that HGNN-DRL outperforms state-of-the-art A2C and TRPO algorithms by 22% and 29%, respectively, in task throughput, while achieving higher security compliance and significantly reduced end-to-end delay.
| Original language | English |
|---|---|
| Pages (from-to) | 11567-11581 |
| Number of pages | 15 |
| Journal | IEEE Transactions on Cognitive Communications and Networking |
| Volume | 12 |
| Early online date | 20 Aug 2026 |
| DOIs | |
| Publication status | Published - 25 Aug 2026 |
Bibliographical note
Copyright © 2026, IEEE. This accepted manuscript version is licensed under the Creative Commons Attribution License CC BY [https://creativecommons.org/licenses/by/4.0/], which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.Funding
This work is supported in part by the National Research and Development Programs of China (2025YFB3109801), National Natural Science Foundation of China (62361011), Guizhou Provincial Science and Technology Plan Project (DXGA[2025]003, DXGA[2025]011,[2026]175,[2026]209), Guizhou Province Scientist Workstation (KXJZ[2025]005), and Guizhou Province Science and Technology Innovation Platform (JSZX(2025)020). (Corresponding authors: S. Zou).
| Funders | Funder number |
|---|---|
| National Key Research and Development Program of China | 2025YFB3109801 |
| [2026]175, DXGA[2025]011, [2026]209, DXGA[2025]003 | |
| KXJZ[2025]005 | |
| JSZX(2025)020 | |
| National Natural Science Foundation of China | 62361011 |
Keywords
- Untrusted networks
- double deep Q-network (DDQN)
- heterogeneous graph neural network (GNN)
Fingerprint
Dive into the research topics of 'Scheduling Over Untrusted 6G Networks: An HGNN-Augmented Reinforcement Learning Approach for Joint Delay-Security Guarantees'. Together they form a unique fingerprint.Cite this
- APA
- Author
- BIBTEX
- Harvard
- Standard
- RIS
- Vancouver