Low-altitude wireless networks (LAWNs) provide agile communication access and edge computing for ground users in hotspots, temporary overload areas, sparse-coverage regions, and emergencies. However, their uncrewed aerial vehicle (UAV) nodes have limited resources and local observations, which restrict cross-region coordination under spatially imbalanced and bursty demand. This article presents an architecture-led framework in which satellites perform slow-timescale wide-area orchestration while UAV clusters retain fast local service execution. A digital twin maintains synchronized network and service states, and the network operation and control center manages model training, validation, and deployment. Three enabling functions form the operational loop. Multimodal and spatiotemporal prediction transforms service, historical, geographic, and event information into regional communication-demand, computing-demand, and urgency maps. A world model completes unreported states, evaluates information freshness, and adapts reporting among high-, medium-, and low-frequency regimes. Hierarchical multiagent reinforcement learning combines predicted demand with freshness-tagged digital-twin states. Satellite-side actors allocate regional backhaul and computing budgets, while parameter-shared UAV actors select local bandwidth-and-offloading profiles. Service outcomes are returned to update the digital twin and subsequent decisions. Controlled simulations under spatially concentrated load show that the hierarchical method reduces penalized average delay by 2.47%–5.45% and improves quality-of-service satisfaction by 0.41–1.68 percentage points relative to flat multi-agent reinforcement learning. The results demonstrate that predictive regional budgeting, adaptive state acquisition, and responsive local execution can jointly improve communication-computing resource management under practical signaling and onboard-inference constraints.
With the evolution toward AI-native 6G networks, large language model (LLM) inference is expected to move from centralized clouds to distributed edge infrastructures embedded in mobile networks. This deployment approach can support low-latency intelligent services using available computation resources at the network edge. Studying distributed LLM inference in mobile networks, however, requires researchers to jointly capture mobile network procedures, communication and computation resource contention, and heterogeneous execution across edge nodes and model partitions. These requirements call for a practical simulation tool that can accelerate research exploration before large-scale physical deployment. Existing network simulators and small-scale testbeds remain limited in capturing the tight coupling between communication and computation in distributed LLM inference over mobile networks. This paper presents CoCoSim, a communication and computation co-simulation platform for distributed LLM inference in mobile networks. Built on the current 5G/5G-Advanced architecture, CoCoSim aims to support research on distributed LLM inference in future AI-native 6G networks. CoCoSim integrates workflow-grounded network modeling, session-driven concurrency simulation, and hardware-aware LLM execution, enabling realistic and scalable exploration of distributed LLM inference designs. Three case studies examine concurrent serving fidelity, wireless-link effects, and communication and computation bottlenecks.
Fluid antenna systems (FASs) have recently attracted increasing attention due to their high flexibility and reconfigurability. By exploiting spatial channel variations through flexible antenna positions, FAS introduces additional spatial degrees of freedom for improving wireless network performance. In this article, we investigate the potential of FAS for anti-jamming communications and discuss how spatial reconfiguration can be used to mitigate hostile jamming. We first present the basic FAS architecture and spatial channel model, and then highlight its advantages over conventional spatial-domain anti-jamming techniques. To capture the hierarchical interaction between a legitimate user and an adversarial jammer, we develop a Stackelberg game-based framework for FAS-aided transmission strategy optimization. Within this framework, a reinforcement learning-based FAS port selection scheme is combined with transmit beamforming to suppress jamming interference. A case study and numerical results are further provided to evaluate the effectiveness of the proposed FAS-aided anti-jamming approach. Simulation results show that the proposed scheme can reduce the outage probability and improve the average achievable rate under jamming attacks.
At present, the main sensing functions supported by the existing Integrated Sensing and Communication (ISAC) systems are the measuring of range, speed and angle. There is relatively little research on the ISAC signal imaging. This paper focuses on the research of ISAC signal design, and especially on the corresponding signal processing algorithms for Synthetic Aperture Radar (SAR) imaging towards low-altitude flight Uncrewed Aerial Vehicle (UAV) scenario (not for satellite scenario). Firstly, facing to the field of UAV low-altitude flight sensing, an ISAC signal supporting the point targets SAR imaging function is designed. The designed ISAC signal is based on the Frequency Modulated Continuous Wave (FMCW) radar signal. The corresponding signal processing algorithm is proposed also. On the one hand, as for sensing, the improved self-focusing imaging algorithm is proposed to process the received target echo signal and achieve the SAR imaging with lower complexity. On the other hand, as for communication, the communication information can be demodulated at the receiver. Secondly, the complexity analysis of the proposed ISAC signal processing algorithm for SAR imaging is provided. The simulation result shows that the complexity of the improved algorithm is lower than the traditional one. Finally, the simulation results show that the designed ISAC signal and its signal processing algorithm not only has the acceptable spectral efficiency (information rate) compared with the traditional communication signals, but also supports the SAR moving point targets imaging function for the low-altitude flight UAV scenario.
The transition to disaggregated 5G and 6G architectures has rendered centralized security models obsolete, necessitating a shift toward decentralized, autonomous defense. While Vertical Federated Learning (VFL) offers a promising path for collaborative detection across the RAN, Edge, and Core without exposing sensitive data, it introduces a critical “trust gap.” Network operators remain hesitant to authorize autonomous mitigation actions driven by “black box” neural networks that lack transparency and accountability. This article proposes a unified framework that bridges this gap by integrating Agentic AI, VFL, and Explainable AI (XAI). We introduce a distributed architecture where agents detect attacks using privacy-preserving VFL, but crucially, leverage a Retrieval-Augmented Generation (RAG) engine to reason before acting. By grounding AI decisions in NIST guidelines and operational Service Level Agreements (SLAs), the system transforms opaque alerts into transparent, evidence-backed mitigation plans. We demonstrate that this approach eliminates AI hallucinations and enables “surgical” responses—such as rate-limiting at the RAN or flow filtering at the Edge—that are fully traceable, policy-compliant, and trustworthy.
