TY  - EJOU
AU  - Alharbi, Abdullah 

TI  - Trustworthy Semantic Knowledge Orchestration for Space-Air-Ground Integrated Edge Computing Using Hierarchical Federated Multi-Agent Deep Reinforcement Learning
T2  - Computers, Materials \& Continua

PY  - 
VL  - 
IS  - 
SN  - 1546-2226

AB  - Space-Air-Ground Integrated Networks (SAGIN) are networks that combine the use of satellite, High-Altitude Platform Systems (HAPS), Unmanned Aerial Vehicles (UAVs), terrestrial base stations, and Internet of Things (IoT) devices to enable ubiquitous networking and distributed edge intelligence. Efficient semantic edge computing is, however, challenging due to resource heterogeneity, as well as the dynamism of the topology and the evolution of semantic information. This study proposes a novel Hierarchical Graph-Constrained Federated Multi-Agent Deep Reinforcement Learning (HGCF-MARL) framework for trustworthy semantic knowledge orchestration in SAGIN. The framework unifies the placement of semantic knowledge bases, task offloading, communication-computation resource allocation, UAV trajectory planning, and coordination between layers. A Semantic Knowledge Freshness-Mismatch Index (SKFMI) is introduced to quantify temporal staleness, distribution drift, and task mismatch, and trust-aware, Byzantine-robust aggregation and event-triggered policy synchronization mechanisms are introduced to enhance reliability and communication efficiency. HGCF-MARL demonstrated a semantic quality of 0.936 and a knowledge-hit ratio of 94.2% across experiments on EuroSAT, CIFAR-10, AG News, and synthetic SAGIN traces. It shows the average reductions of 13.8% and 56.0% in total energy consumption and federated signaling volume, respectively, compared with the best corresponding baselines, and achieves up to 99.2% model-replacement attack detection rate while maintaining up to 94.1% task-completion ratio in the 10% malicious-participant setting. These results show the effectiveness, robustness, and scalability of HGCF-MARL for dynamic SAGIN edge intelligence.
KW  - Space-Air-Ground Integrated Networks; semantic edge computing; semantic knowledge orchestration; multi-agent deep reinforcement learning; federated learning; Byzantine robustness

DO  - 10.32604/cmc.2026.089052