Federated learning (FL) has emerged as a promising paradigm for privacy-preserving model training across distributed edge devices, enabling local data utilization without explicit sharing. However, in edge computing environments characterized by heterogeneous resources and intermittent connectivity, FL remains vulnerable to gradient leakage attacks (GLA), where adversaries reconstruct private data from shared model updates. Although existing defenses, such as differential privacy (DP) and gradient compression, offer partial mitigation, they often result in significant performance degradation or an increased communication overhead. In this paper, we analyze the risk of privacy leakage that is highly sensitive to the client-side training configurations and gradient magnitudes. Based on this, we propose a risk-aware FL framework tailored for the edge scenarios, which not only performs per-device privacy risk assessment but also introduces subtractive dithering quantization to the inject controllable Gaussian noise into local models. Additionally, a noise-aware aggregation strategy is presented by adjusting each client's contribution to preserve the global model utility. Experimental results on FashionMNIST and CIFAR-10 demonstrate that the proposed framework achieves strong defense against the GLA, reduces the communication costs by over 50%, and maintains a competitive accuracy.

Fluid Antennas (FAs)-assisted Unmanned Aerial Vehicle (UAV) networks leverage the FA position adaptivity and flexible beamforming to overcome the limitations of Fixed-Positioned Antennas (FPAs) in dynamic UAV channels and Multi-User (MU) interference. This letter investigates a dual FA-assisted UAV network for MU-Multiple-Input-Multiple-Output (MIMO) downlink communications, aiming to maximize the average achievable rate through the joint optimization of UAV trajectory, the transmit/receive FA positions, and beamforming. The formulated problem is highly coupled and non-convex. Accordingly, an efficient Alternating Optimization (AO)-based algorithm is developed for decomposed subproblems, yielding a suboptimal solution. Numerical results demonstrate significant performance gains of 120

Designing receivers for integrated sensing and communications (ISAC) systems is challenging due to mutual interference between communication and sensing. To address this, we first propose a basic MMSE receiver for joint processing and then introduce two enhanced and low-complexity designs, a fixed-weight MMSE receiver and an adaptive-weight MMSE receiver. The fixed-weight MMSE receiver provides controlled performance trade-offs through weight factors, while the adaptive-weight MMSE receiver dynamically optimizes weights based on instantaneous channel conditions, enhancing system flexibility. We develop comprehensive theoretical analyses for both designs, establishing performance bounds for the fixed-weight approach and convergence properties for the adaptive mechanism. Simulation results demonstrate that the proposed MMSE-based receivers achieve superior robustness against sensing power variations and significantly reduce complexity compared to conventional methods, while maintaining balanced performance between communication and sensing functionalities.

Large AI models are reshaping intelligent manufacturing from isolated automation toward knowledge-intensive, model-assisted production systems. Yet their industrial value depends not on model scale alone, but on how language, vision, code, sensor data, engineering knowledge, and feedback mechanisms are integrated into deployable manufacturing workflows. This review examines recent progress in large AI models for intelligent manufacturing, covering model architectures, adaptation strategies, system integration, and applications across product development, production processes, equipment maintenance, and manufacturing services. A lifecycle-based framework is used to organize the literature and distinguish model capabilities from the data resources, retrieval mechanisms, simulation and optimization tools, digital twins, edge-cloud infrastructure, and human validation required for deployment. Current evidence suggests that large models show more reliable value in bounded, information-rich tasks, whereas safety-critical control and production-scale autonomy remain insufficiently validated. The review further summarizes challenges in data quality, domain adaptation, reliability, interpretability, latency, cybersecurity, cost, benchmarking, and responsibility allocation. By linking application scenarios, system-level enablers, and evidence maturity, this review provides a structured perspective for assessing the practical value of large AI models in manufacturing.

Information age and latency are of both vital importance for providing interactive system operation and immersive user experience. In this paper, we wrestle with the challenge of balancing information age and latency in mobile edge computing (MEC) networks with short packet communications. Specifically, we derive analytical expressions for the average peak age of information (PAoI) and average latency in terms of the blocklength, offloading ratio, and maximum allowable transmission times. In order to strike a performance tradeoff, we devise an algorithm to minimize the weighted sum of average PAoI and average latency. Simulation results demonstrate that our solution aligns with the PAoI-latency Pareto front, matching the performance of high-complexity exhaustive search while outperforming comparative baseline methods across diverse local and edge computing frequency.

On-device inference offers privacy, offline use, and instant response, but consumer hardware restricts large language models (LLMs) to low throughput and capability. To overcome this challenge, we present prima.cpp, a distributed on-device inference system that runs 30-70B LLMs on consumer home clusters with mixed CPUs/GPUs, insufficient RAM/VRAM, slow disks, Wi-Fi links, and heterogeneous OSs. We introduce pipelined-ring parallelism (PRP) to overlap disk I/O with compute and communication, and address the prefetch-release conflict in mmap-based offloading. We further propose Halda, a heterogeneity-aware scheduler that co-optimizes per-device CPU/GPU workloads and device selection under RAM/VRAM constraints. On four consumer home devices, a 70B model reaches 674 ms/token TPOT with <6% memory pressure, and a 32B model with speculative decoding achieves 26 tokens/s. Compared with llama.cpp, exo, and dllama, our proposed prima.cpp achieves 5-17× lower TPOT, supports fine-grained model sizes from 8B to 70B, ensures broader cross-OS and quantization compatibility, and remains OOM-free, while also being Wi-Fi tolerant, privacy-preserving, and hardware-independent. The code is available at https://anonymous.4open.science/r/prima-cpp.

The advent of Large Multimodal Models (LMMs) offers a promising technology to tackle the limitations of modular design in autonomous driving, which often falters in open-world scenarios requiring sustained environmental understanding and logical reasoning. Besides, embodied artificial intelligence facilitates policy optimization through closed-loop interactions to achieve the continuous learning capability, thereby advancing autonomous driving toward embodied intelligent (El) driving. However, such capability will be constrained by relying solely on LMMs to enhance EI driving without joint decision-making. This article introduces a novel semantics and policy dual-driven hybrid decision framework to tackle this challenge, ensuring continuous learning and joint decision. The framework merges LMMs for semantic understanding and cognitive representation, and deep reinforcement learning (DRL) for real-time policy optimization. We starts by introducing the foundational principles of EI driving and LMMs. Moreover, we examine the emerging opportunities this framework enables, encompassing potential benefits and representative use cases. A case study is conducted experimentally to validate the performance superiority of our framework in completing lane-change planning task. Finally, several future research directions to empower EI driving are identified to guide subsequent work.

Future wireless networks are expected to increasingly rely on communication-efficient distributed learning and optimization to coordinate decisions among spatially separated agents under limited signaling resources. In 6G fully decoupled radio access networks (FD-RANs), where uplink (UL) and downlink (DL) base stations (BSs) operate independently and support the softwarization scheme, coverage flexibility and spectrum efficiency can be improved. Nevertheless, asymmetric interference, discrete resource coupling, and heterogeneous BS capacities pose significant challenges to communication-efficient coordination. To address these challenges, this paper proposes a fairness-aware joint UL-DL association and resource allocation framework with a dual-layer fairness mechanism based on congestion pricing and Jain's fairness index. Since centralized joint optimization leads to a large action space, while independent UL/DL optimization may weaken cross-direction fairness, the problem is reformulated as a distributed two-layer multi-armed bandit (MAB) model. In this model, UL and DL decisions are learned in separate layers, while fairness and pricing signals provide coupling across the two directions. Based on this formulation, a distributed two-layer fairness-driven bandit algorithm is developed, where mean-index exploration, adaptive exploitation intervals, and interrupt-driven coordination jointly reduce signaling overhead and support fairness convergence. Theoretical analysis establishes logarithmic regret and convergence to equilibrium. Simulation results against recent FD-RAN baselines show that the proposed framework achieves low outage probability, maintains a high Jain's index, and outperforms existing schemes in both fairness and efficiency under realistic FD-RAN scenarios.

Driven by the increasing deployment of Internet of Things (IoT) devices and significant progress in Uncrewed Aerial Vehicle (UAV) technologies, UAV-assisted Mobile Edge Computing (UMEC) has emerged as an effective approach to address the challenges of constrained ground network coverage and insufficient computational resources. Leveraging their mobility, flexible deployment, and low cost, UAVs can dynamically support edge networks by providing computation and communication services for latency and energy sensitive tasks. Nevertheless, task scheduling and UAV trajectory optimization in UMEC systems still face significant challenges. To address the task scheduling problem for energy-sensitive tasks in the UMEC system, this paper first formulates a mathematical model of the system's energy consumption and establishes an optimization objective aimed to optimize the energy cost of UMEC system. Building upon this, we propose a task scheduling algorithm that integrates Multi-Agent Proximal Policy Optimization (MAPPO) with a bargaining game-based resource coordination mechanism. The proposed algorithm reduces the overall energy consumption and enhances the system performance through a bargaining-based resource allocation and matching strategy, alongside MAPPO-driven trajectory optimization. Simulation results show that our method outperforms the baseline algorithms in terms of system utility, achieving up to a 7.4% improvement.

A Reconfigurable Intelligent Surface (RIS) enhances the performance of wireless networks by smartly reconfiguring the wireless propagation environment. This enhancement is highly dependent on the number of RIS elements, but practical challenges like high channel acquisition overhead and power consumption limit the scalability of traditional RIS systems. Recently, proposed irregular RIS has been shown to tackle these challenges effectively. However, its potential advantages are not fully realized since the topology and precoding designs are alternatively optimized, which does not consider the impact of topology changes on precoding. In this paper, we leverage the benefits of irregular RIS and propose new approaches to fully exploit its potential. The key idea is to enhance the system capacity by enabling simultaneous optimization of the irregular RIS topology and precoding design. Specifically, we propose the standard particle swarm optimization (PSO) algorithm to jointly optimize the RIS topology, beamforming, and reflection coefficients. Unlike the conventional approach, the PSO simultaneously evaluates potential solutions via particle updates that enhance the system capacity significantly. Next, we introduce an improved variant of the PSO designed to improve the exploration-exploitation tradeoff and speed up the convergence. Specifically, the improved PSO approach enables the dynamic adjustments of PSO parameters throughout the process, reducing the risk of premature convergence and significantly improving the overall performance compared to the standard PSO. The effectiveness of the proposed improved PSO is validated through a complex multi-modal Michalewicz function benchmark test, which illustrates that the obtained results are virtually the same as the expected results for different dimensional spaces. Besides, we also provide convergence and complexity analysis of the improved PSO. Numerical results and analysis demonstrate that the proposed algorithms significantly enhance the system capacity with low complexity by simultaneously optimizing the topology and precoding design.

Human activity prediction (HAP) is crucial for enabling intelligent smart home services; yet, it is often hindered by the scarcity of high-quality, multidimensional datasets. Existing datasets are typically fragmented, capturing either long-term activity sequences or short-term device interactions, but rarely both in a unified manner. Traditional data collection methods are costly and time-consuming, while conventional simulation techniques struggle to generate diverse and logically coherent behavior sequences. To address these limitations, we propose SmartLLM, a novel large language model (LLM)-based simulation framework for automated generation of multidimensional smart home datasets. SmartLLM simulates simulated agents with distinct profiles (e.g., old man, remote worker, and holiday maker) performing daily activities within configurable home environments, generating temporally aligned sequences across activity-device-sensor dimensions. We generate two months of simulated data for three user profiles and validated their plausibility through activity distribution visualization, statistical perplexity analysis, and case studies. Multidimensional feature validation experiments further demonstrate that our multidimensional data significantly enhances the accuracy of activity prediction models compared to using single-dimensional features. This work successfully addresses key bottlenecks in smart home data acquisition and provides a scalable, high-quality data foundation for advancing smart home algorithm research.

The sixth-generation (6G) wireless communications aim to achieve wide coverage, ubiquitous connectivity, and ultra-reliable low-latency transmission by unleashing the potentials of cross-domain collaboration in space, semantics, and polarization. Recently, stacked intelligent metasurface (SIM) has emerged as a revolutionary paradigm for enabling multi-domain communication functionalities, i.e., offering high-degree-of-freedom (DoF) spatial beamforming, parallel light-speed semantic computing, and flexible polarization manipulation with low hardware cost and high scalability. However, the existing SIM-related works merely focus on single-domain manipulation, while the potentials of SIM for cross-domain collaboration remain unexploited in 6G wireless networks. To this end, we propose a cross-domain collaborative SIM (C-SIM), where the above capabilities in spatial, semantic, and polarized domains are seamlessly integrated to promotes a significant performance gain for 6G wireless networks. This paper provides a comprehensive overview of C-SIM, encompassing its preliminaries, architecture, and potential benefits. Moreover, we discuss the cutting-edge applications and identify open challenge associated with enabling C-SIM. Finally, a case study and numerical results are presented to quantify the benefits of C-SIM for 6G wireless networks.

Graph anomaly detection (GAD) aims to identify irregular nodes or structures in attributed graphs. Neighbor information, which reflects both structural connectivity and attribute consistency with surrounding nodes, is essential for distinguishing anomalies from normal patterns. Although recent graph neural network (GNN)-based methods incorporate such information through message passing, they often fail to explicitly model its effect or interaction with attributes, limiting detection performance. This work introduces NeiGAD, a novel plug-and-play module that captures neighbor information through spectral graph analysis. Theoretical insights demonstrate that eigenvectors of the adjacency matrix encode local neighbor interactions and progressively amplify anomaly signals. Based on this, NeiGAD selects a compact set of eigenvectors to construct efficient and discriminative representations. Experiments on eight real-world datasets show that NeiGAD consistently improves detection accuracy and outperforms state-of-the-art GAD methods. These results demonstrate the importance of explicit neighbor modeling and the effectiveness of spectral analysis in anomaly detection. Code is available at: https://github.com/huafeihuang/NeiGAD.

The Payment Channel Network (PCN) is an effective Layer-2 solution for addressing the scalability issues of blockchain-based cryptocurrencies. However, in practice, payment channels often suffer from balance depletion on one side, preventing further transactions from being executed through that channel. Rebalancing methods can alleviate this issue by actively shifting available balances across different payment channels to replenish depleted ones at minimal cost. However, existing rebalancing methods often determine the target balance of each channel subjectively, without ensuring that the rebalancing operation effectively increases the total value of executable transactions or maximizes balance-shifting efficiency within the PCN. To address these limitations, this paper establishes rebalancing evaluation models to quantify the impact of individual rebalancing operations on the total executable transaction value within the PCN. Based on these models, we design a hybrid rebalancing method that adaptively adjusts the rebalancing operation type, target balance, and participant selection according to the available balance and transaction workload of each payment channel, thereby maximizing the overall benefit of rebalancing operations on the total executed transaction value of the PCN. Experimental results demonstrate that our method improves the transaction success rate by up to 34.7% and reduces the average transaction latency by up to 4.4% compared with existing methods.

Event recognition in unmanned aerial vehicles (UAVs)-based monitoring systems is crucial for geospatial analysis, environmental surveillance, and disaster response. However, federated learning (FL)-based approaches for real-time event recognition face significant challenges, including spatiotemporal heterogeneity, constrained computational resources, and communication limitations in distributed UAV networks. Although FL enables decentralized machine learning with enhanced data privacy, its application to temporal event recognition in UAV-based remote sensing (RS) systems requires addressing data heterogeneity and maintaining spatiotemporal coherence across distributed nodes. To address these challenges, we propose a split FL (SFL) framework tailored for UAV-assisted event recognition in geospatial monitoring. The proposed SFL architecture partitions computational workloads between UAV-based edge clients and a central server, optimizing both on-device efficiency and global model performance. At the client level, lightweight convolutional models extract spatiotemporal features from UAV-captured video sequences, reducing computational complexity and transmission overhead. These extracted features are transmitted to a central server, where higher order temporal dependencies are learned, enabling robust event classification. To enhance adaptability in dynamic environments, we integrate dynamic video chunking, adaptive temporal pooling, and modality-agnostic feature aggregation, ensuring efficient processing of variable-length sequences while minimizing bandwidth constraints. Experimental evaluations on standard UAV-based geospatial datasets demonstrate that the proposed SFL framework significantly outperforms conventional FL approaches in terms of classification accuracy, communication efficiency, and scalability. This work provides a scalable, privacy-preserving, and computationally efficient solution for real-time temporal event recognition in UAV-assisted geospatial monitoring applications.

Behavior trees have found wider applications in various fields, including task and network planning. However, how to generate the behavior tree according to specific intents for diverse scenarios remains a challenging problem. Manual approaches offer low efficiency, high error rates, and a heavy reliance on domain-specific knowledge. Automated methods often struggle to accurately interpret complex intents. With the emergence of large language models, research on generating behavior trees from natural language intents has gained significant attention. This article presents a generic behavior tree generator framework using large language models. Based on the framework, we propose a generic behavior tree automatic generator scheme. We verify the effectiveness of the proposed scheme through a satellite task and network planning use case. Experimental results show that compared with traditional generation methods based solely on pre-trained large language models, the proposed method improves behavior tree generation time and accuracy by 70.01% and 45.67%, respectively.

Enhanced by inter-satellite links and satellite direct-to-device capabilities, satellite networks can offer low-latency communication globally. However, limited spectrum resources and the capacity bounds of the Shannon's information theory pose fundamental challenges for supporting bandwidth-intensive multimedia services. Semantic communication (SemCom) offers a promising solution by transmitting compressed semantic representations instead of raw data, thereby alleviating bandwidth pressure. However, it also introduces SemCom-related constraints that render conventional schemes such as contact graph routing inapplicable. To overcome this challenge, we investigate SemCom-compliant path selection and formulate it as a non-NP hard mixed-integer linear programming problem. To address the problem, we develop a graph-based scheme that exploits the special structure of the solution space, the sparsity of SemCom-capable satellites, and the property of Dijkstra's algorithm, thus achieving optimal solutions with polynomial-time complexity. Simulation results on the Starlink constellation confirm that the proposed scheme facilitates SemCom with negligible computational overhead and significant bandwidth reduction. While the bandwidth reduction comes at the cost of increased delay and path hops, these effects are shown to be mitigatable through higher SemCom deployment in a satellite network or by enabling semantic processing at the user side.

The Internet of Things is constrained by energy consumption and battery limitations, leading to high network maintenance costs and environmental impact. Battery-free ambient backscatter communication (AmBC) provides a sustainable alternative by enabling passive tags to harvest RF energy for continuous data transmission. The inherently low power of AmBC signals increases signal outage probability and bit error rates (BER), requiring effective adaptation mechanisms. This paper proposes a new reflection coefficient adjustment approach that enables both energy harvesting and reliable data transmission at each passive tag. In our decentralized multi-agent framework, tags compete to achieve higher transmit powers for reliable communication. To mitigate interference between tags, we model the interaction among tags as a non-cooperative Cournot game, where each tag acts as a rational agent adjusting its reflection coefficient to maximize its own utility—defined as a trade-off between transmission reliability and interference mitigation. Our game-theoretic approach eliminates the need for explicit coordination or information exchange, promoting scalability and robustness. Numerical results show that our approach significantly outperforms existing methods by reducing BER and improving energy efficiency, making it a viable solution for sustainable, battery-free IoT communications.

The proliferation of connected and autonomous vehicles within the Internet of Vehicles (IoV) is generating a data deluge, straining current wireless networks which primarily focus on raw data transmission. This creates a semantic gap between data volume and actionable intelligence, hindering the realization of truly intelligent, efficient, and reliable vehicular services for the sixth generation of mobile networks (6G). To bridge this gap, this article proposes a novel architectural vision: Intelligent Semantic Agents empowered IoV (ISA-IoV). ISA-IoV deploys a hierarchy of Semantic Agents, software entities powered by foundation models, across cloud, edge, and end devices to collaboratively perform semantic perception, reasoning, and communication. We define the scalability and adaptability requirements for ISA-IoV and introduce three unique Key Performance Indicators to guide its development. The article further surveys key enabling technologies, discusses major implementation challenges, and outlines promising future directions. By integrating agentic artificial intelligence with semantic communication over a Space-Air-Ground Integrated Network, ISA-IoV presents a concrete roadmap towards semantic-native and intelligence-everywhere 6G vehicular networks.

Space-Air-Ground Integrated Networks (SAGINs) constitute a large-scale heterogeneous network that integrates satellite, aerial, and terrestrial segments with highly dynamic connectivity and diverse link characteristics. Such time-varying system conditions make sequential routing decisions challenging for maintaining Quality of Service (QoS) when topology, link quality, and node resources change rapidly. To address this challenge, this paper proposes a deep reinforcement learning-based multi-objective QoS-aware routing framework named Dual-Critic Multi-Objective Routing (DC-MOR). DC-MOR jointly optimizes bandwidth utilization, resource utilization, energy consumption, and end-to-end delay, and improves routing stability and decision robustness through a dual-critic evaluation structure. In addition, DC-MOR incorporates mobility-aware state representations, a dynamically weighted multi-objective reward function, and a topology-constrained action mapping mechanism to produce feasible routing decisions under time-varying connectivity. Experimental results demonstrate that DC-MOR outperforms DDPG by 14.3% in bandwidth utilization and 17.8% in resource utilization, achieves 2.4% lower energy consumption, and reduces end-to-end delay by 7.9%. These results indicate that DC-MOR provides an efficient and robust routing framework for large-scale dynamic SAGINs.