Hyperledger Fabric is a popular permissioned blockchain platform. Although many existing works on the overall performance of Hyperledger Fabric are available, a decomposition of each phase in Hyperledger Fabric remains to be explored. While the overall performance of Hyperledger Fabric might provide an end-user with sufficient performance information when processing a transaction, it is far from informative when deploying a distributed system with specific performance goals, except for understanding each decomposition phase in Hyperledger Fabric. This paper presents a client-driven measurement framework to characterize the transactions of each decomposition phase in Hyperledger Fabric, based on the Fabric SDK Node.js. We evaluate the Hyperledger Fabric performance and have some interesting observations, e.g., 1) The number of CPU cores linearly impacts the throughput of an endorsing peer. 2) Raft-based ordering service shows good scalability with the number of ordering service nodes. 3) Waiting time dominates a transaction's latency in the order phase when encountering unfit block-setting parameters. We then identify the dominant resource in each decomposition phase, propose a stochastic computation model and an alpha-beta communication model to analyze the performance of each decomposition phase. Finally, we validate the performance model's accuracy in two clusters. And we propose a strategy to reduce waiting time when it dominates transaction latency in the order phase under unfit block-setting parameters, by automatically adjusting block-setting parameters, achieving at least 20% reductions in waiting time.
With the widespread adoption of interactive, mobile, and wearable technologies, an increasing amount of privacy-sensitive and statistically heterogeneous sensor data is being continuously generated by ubiquitous user devices. Federated learning (FL) enables multiple parties to jointly train a machine-learning model while keeping their raw data private. Nonetheless, under realistic conditions characterized by pronounced statistical heterogeneity across client distributions, traditional FL methods suffer from degraded performance. Moreover, uniform privacy mechanisms fail to accommodate the diverse privacy needs of individual users. To this end, we propose a privacy-preserving FL framework, FedPriv, for heterogeneous mobile sensor data, aiming to improve the performance of the model in a heterogeneous data environment while protecting user privacy. Specifically, FedPriv is built upon two core components that we design: (1) a novel aggregation mechanism that dynamically adjusts each client's contribution according to the discrepancy between local and global data distributions, thereby mitigating adverse effects of data heterogeneity on the model performance; and (2) a personalized differential privacy strategy that allocates privacy budgets adaptively, depending on each client's data sensitivity, enabling fine-grained privacy protection without compromising utility. In addition, FedPriv can be seamlessly integrated as a plug-and-play module into existing FL pipelines, enhancing model performance without requiring architectural modifications. Security analysis demonstrates that FedPriv satisfies formal $\epsilon$-differential privacy guarantees while supporting personalized privacy levels across users. Extensive experiments on six real-world datasets, consisting of four image datasets and two human activity recognition datasets, show that FedPriv achieves a 2.75%-5% improvement in model performance and reduces convergence time by an average of 15% when compared to existing baselines under heterogeneous data settings.
Uncrewed aerial vehicle (UAV) swarms are expected to become a key component of future low-altitude wireless networks (LAWNs), enabling agile communication, sensing, and distributed computing services. However, the highly dynamic network topology and the strong coupling among UAV mobility, wireless interference, task execution, and energy consumption make proactive and long-horizon control extremely challenging. To address this problem, we propose a proactive control framework that fuses digital twin (DT) and world model (WM) for dynamic low-altitude UAV swarm networks. The DT provides a physics-consistent virtual replica of the UAV swarm network, capturing UAV dynamics, wireless communication, task queues, and energy evolution. In contrast, the learned WM focuses on predicting control-induced mobility dynamics in a compact latent space. By integrating the WM into a model predictive control (MPC) loop, the proposed framework enables efficient multi-step lookahead optimization of joint UAV mobility control, communication scheduling, and task allocation decisions. Based on this framework, we formulate the proactive control problem as a finite-horizon stochastic optimization that maximizes long-term network utility subject to constraints on UAV kinematics, collision avoidance, communication reliability, energy sustainability, and task deadlines. The formulated problem is solved using the proposed DT-WM-MPC framework, where the WM performs efficient mobility prediction and the DT completes coupled network states and verifies candidate decisions against physical and operational constraints. Extensive experiments demonstrate that the proposed DT-WM-MPC framework consistently outperforms reactive and non-predictive baselines in terms of task completion and cumulative system utility, while reducing planning latency compared with full-DT-based MPC. Furthermore, the learned WM achieves accurate multi-step rollout prediction and maintains stable performance across different UAV swarm scales. The results highlight the potential of integrating DTs with WMs to enable scalable, safe, and proactive control for intelligent LAWNs.
Low-altitude internet of things (IoT) networks impose stringent requirements on beam alignment due to highly dynamic environments and limited onboard resources. Conventional beam training approaches incur substantial overhead and exhibit degraded robustness under such conditions. To achieve reliable beam alignment with minimal resource consumption, this paper proposes an edge intelligence framework that integrates sensing, communication, and computation (ISCC). Specifically, a federated learning mechanism is employed to collaboratively fuse local sensor knowledge across multiple nodes. To enable beam training, integrated sensing and communication (ISAC) is leveraged as a self-supervised sensor interface to generate labeled sensing dataset directly at distributed edge nodes. Over the-air computation (AirComp) is then adopted for local model aggregation over shared spectrum, supported by an ISAC-based dynamic transmission control scheme. Apart from designing a resource-efficient system, a joint optimization of sensing, communication and computation resources is formulated to minimize the overall energy consumption. By deriving FL convergence bound, the optimization problem is solved via Sequential Quadratic Programming algorithm. Simulation results show that the proposed system significantly improves beam alignment accuracy in dynamic environments while reducing energy and spectrum consumption.
This paper investigates the robust tracking control problem of a networked commercial aircraft system during the ascent phase under communication constraints. The aircraft dynamics and the controller are interconnected through a communication network, where efficient utilization of communication resources is critical for maintaining tracking performance. To address unknown external disturbances and model uncertainties, an adaptive sliding mode scheme is developed to estimate disturbance effects online without requiring prior knowledge of disturbance bounds. A backstepping-based robust control framework is then constructed to guarantee longitudinal tracking stability of the networked aircraft system. To reduce communication burden while preserving closed-loop performance, a switched event-triggered communication mechanism is proposed, where the triggering policy is autonomously adjusted according to the evolution of system stability. The proposed strategy establishes a tradeoff between tracking performance and network resource utilization by significantly reducing transmission events. Rigorous analysis establishes the boundedness of all closed-loop signals, guarantees tracking convergence, and excludes Zeno behavior. The effectiveness of the proposed network-aware control framework is demonstrated through simulations on a Boeing 747 aircraft model during the ascent phase.
The advancement of Internet of Vehicles (IoV) has led to significant progress in integrating Artificial Intelligence Generated Content (AIGC) into vehicular networks. However, current research predominantly focuses on traditional performance metrics such as latency and energy consumption, while overlooking the joint optimization of semantic consistency, image quality, and communication metrics, which affect the QoS of users directly. In this paper, we propose a novel Aerial-Ground Cooperative Diffusion (AG-CoDiff) system in an Unmanned Aerial Vehicle (UAV)-assisted IoV, which integrates Collaborative Fine-Tuning (CoFT) and Collaborative Inference (CoInf) for a complex scene understanding in autonomous driving. In the CoFT phase, vehicles capture image data and transmit it to UAVs for filtering and fine-tuning, while in the CoInf phase, UAVs perform shared inference, and vehicles conduct local inference based on received intermediate results. A novel satisfaction metric is introduced to balance inference latency, content quality, and energy consumption simultaneously. Additionally, a dynamic resource allocation algorithm based on Twin Delayed Deep Deterministic Policy Gradient (TD3) and an inference partitioning strategy using the Dream Optimization Algorithm (DOA) are proposed. Experimental results validate the effectiveness of the proposed system. In the CoFT phase, the proposed scheme improves the image quality score by 82.87% and 104.55% over the random-cooperative and non-cooperative baselines, respectively, and improves CLIPScore by 2.29% and 7.50%. In the CoInf phase, the DOA-based partitioning strategy achieves 15.33%–33.64% improvement in objective value over competing heuristics, demonstrating the effectiveness of AG-CoDiff in jointly balancing latency, generation quality, and energy consumption.
Federated Learning (FL) has gained significant attention for its ability to collaboratively train machine learning models across distributed clients. However, traditional FL approaches often overlook the issues of both participation willingness of clients and their early dropouts due to energy depletions over a long-term training. This will result in insufficient training data in FL, and further weaken model generalizations. In this paper, by considering both energy limits and fairness selection of clients, we propose an iterative Auction and Reinforcement Learning (RL) enabled sustainable FL, i.e., SFARL, in which a sustainable RL, i.e., FSRL, is designed to satisfy clients' energy limits by determining the optimal local training iterations. To promote fairness among clients, a client selection algorithm and a two-part payment rule are further designed. The theoretical analysis demonstrates that SFARL upholds incentive compatibility, individual rationality, and fairness. Experimental results show that SFARL consistently achieves top-tier accuracy on MNIST, Fashion-MNIST, CIFAR-10, and CIFAR-100, while preserving the largest number of surviving clients and the lowest fairness metric among the compared methods. In addition, SFARL exhibits the best energy utilization efficiency and the one-time reporting mechanism significantly reduces the overall system consumption.
The rapid growth of edge cloud infrastructure introduces new challenges in managing energy consumption data while ensuring privacy during energy trading. Existing studies focused on smart metering privacy, local electricity markets, and privacy-aware control, whereas privacy-preserving cloudlet energy trading with battery support remains largely unexplored. This paper addresses these challenges by proposing an approach that effectively motivates cloudlets to trade their superfluous energy to minimize energy cost while protecting the privacy of energy usage information and the corresponding underlying data processing patterns. By applying Lyapunov optimization in coupling with differential privacy (DP), we develop a privacy preserving energy trading algorithm where the cloudlet battery level and data backlog are modeled as virtual queues. This trans forms the problem of interest from a long-term average objective to a convex per-instant optimization problem solved effectively. We analytically establish the trade-off between minimizing energy costs and maintaining the stability of both data queues and batteries. It is revealed that the effect of DP on average system costs can be minimized by allowing higher data queue and battery levels, ensuring robust privacy protection without sacrificing cost efficiency. The minimum capacity requirement for the batteries is identified for privacy guarantee. Simulations confirm the effectiveness of our approach, demonstrating up to a 34% reduction in average energy costs for cloudlets.
Intelligent edge networks are increasingly powered by collaborative agentic AI systems, where edge devices operate as autonomous agents. However, non-independent and identically distributed (Non-IID) edge data introduces severe statistical heterogeneity, causing local models to deviate from the global consensus and undermining the trustworthiness of agent updates. In this work, we systematically study this problem under three representative multi-level data imbalance scenarios: i) local imbalance and global balance, ii) local balance and global imbalance, and iii) dual-imbalance. Under such conditions, agents often overfit their skewed local data while forgetting previously learned global knowledge, leading to unstable collaboration. To address this challenge, we propose Trustworthy Agentic Federated Learning (TAFL), a dual-constraint framework that balances autonomous local adaptation with trustworthy global learning. The framework employs a two-stage optimization strategy: a prototype-based semantic alignment stage that establishes a shared feature space with minimal communication overhead, followed by a trustworthy consolidation stage that anchors globally important parameters to prevent the forgetting of global knowledge. Extensive experiments demonstrate that TAFL consistently outperforms state-of-the-art methods, achieving higher accuracy, faster convergence, and more stable collaboration under the three imbalance scenarios.
The modeling of associative memory increasingly draws on concepts from network science. While Modern Hopfield Networks (MHNs), often viewed as densely connected graphs with attention mechanisms, achieve high performance, their computational complexity imposes scalability challenges due to their computational cost. This paper introduces and analyzes Structural Associative Memories (SAMs), a new class of memory models based on sparse graph topologies. We provide a comparative analysis of SAMs and MHNs, examining how network structure affects storage capacity, retrieval accuracy, and efficiency. Theoretical results show that SAM capacity can scale quadratically with the number of nodes, offering clear advantages in large-scale systems. Experiments on scene recognition validate these findings and further demonstrate how mechanisms such as inhibition and parameters like context size shape performance. The results indicate that, while MHNs remain effective in dense data settings, SAMs deliver a computationally efficient and scalable alternative for massive, sparse knowledge-graph memories. This work advances network science by presenting a scalable, graph-based architecture for cognitive modeling and quantifying the relationship between network topology and associative function.
We study radio node (RN) placement for indoor enterprise networks. Using stochastic geometry (SG), we derive the meta-distribution (MD) of the SINR for a test user equipment (UE), with and without cooperation from outdoor macro base stations (MBSs), and compare these results with an integer linear programming (ILP) approach. SG provides an estimate of the required number of RNs but not their locations, while ILP can yield inaccurate local optima and requires high computational power. To address this, we investigate clustering-based algorithms for initializing RN locations using UE location distributions. Along with standard methods, we propose a weighted k-harmonic means (WKHM) clustering strategy tailored to maximize SINR. We then introduce a constrained sequential minimum cut algorithm, , to merge multiple RNs into larger cells and further improve SINR. This is the first work that integrates SG-based statistical analysis, optimization, and clustering to obtain system design insights, dimensioning rules, and planning strategies for enterprise 5G.
In the realm of smart manufacturing and the Industrial Internet of Things (IIoT), robust and covert signal transmission is critical to maintaining secure automation workflows and defending against cyber threats. Traditional signal embedding schemes are rule-based and static by introducing randomness to mask the existence of signal transmission, thus facing the issues of environmental sensitivity, insufficient adaptability, etc. As a highly regarded method featuring dynamic signal covert, Generative Adversarial Network (GAN) provides a novel solution for signal embedding by dynamic games between the corresponding Generator and Discriminator therein. However, GAN-based signal embedding still faces challenges, including environmental sensitivity and insufficient adaptability, in particular in resource-constrained edge settings. Motivated by this, this paper proposes an innovative compression-aware steganographic framework empowered by agentic GAN, owning the capacity of autonomous signal embedding and extraction in resource-constrained edge environments. With semantic-aware techniques, the system covertly fuses sensitive industrial control signals and sensor streams into machine vision imagery, upholding imperceptibility and fidelity while emulating edge-specific distortions like JPEG compression, Gaussian noise, and resolution scaling. Empirical assessments reveal exceptional transparency (PSNR/SSIM), payload capacity, and distortion resilience, affirming its prowess in delivering low-latency, secure data safeguarding within autonomous IIoT frameworks.
As cyber threats continue to evolve, there is a need for Autonomous Cyber Defense (ACD) strategies capable of fast and context-aware responses. Reinforcement learning (RL) has shown promise in automating cyber defense by exploring and learning effective countermeasures. However, RL often struggles with sparse reward signals and insufficient context to handle diverse attack scenarios. Furthermore, the convergence time of an RL agent is often high, making it difficult to train the agent in online settings. To address these challenges, we propose a large language model (LLM)-enhanced RL method that builds and queries a knowledge base (KB) derived from agent-environment interactions. We leverage the pre-trained knowledge of an LLM on different cybersecurity frameworks and use the LLM to analyze parts of the KB to generate appropriate actions for the RL agent. The LLM-generated output is infused into the RL training process to improve performance and reduce convergence time. To validate our approach, we formulate two RL problems: a contextual bandit problem, which accounts for possible misclassifications of network flows by the detection module, and a multi-step RL problem, which considers that adversarial actions may be missed by monitoring or detection tools. For the contextual bandit problem, we develop a custom environment guided by the MITRE ATT&CK framework, while for the multi-step RL problem, we use a prominent Cybersecurity simulation platform named CybORG. Experimental results show that our proposed approach outperforms the baseline RL by over 75% and 65% in the contextual bandit and multi-step RL settings, respectively, in terms of selecting more effective actions.
The stability and energy efficiency of power grids are fundamentally governed by their internal power flow behaviors, yet a unifying principle bridging network flow and system dynamics has been lacking. This work establishes a clear causal chain: power flow distribution determines total line power, which in turn governs both energy dissipation and synchronizability. Specifically, under fixed total energy demand, minimizing total line power directly enhances synchronizability, allowing grid optimization to be reformulated as a minimum network flow problem. Based on this insight, three strategies aimed at minimizing total line power are proposed: load shedding, generation allocation, and topology optimization. However, such minimization inherently comes at the cost of reduced controllability. To compensate for this loss, a flow-driven algorithm for optimal driver placement and a corresponding distributed controller are developed to jointly reduce control energy and enhance control lability. Collectively, this work offers a practical framework for constructing more sustainable, resilient, and efficient power grids.
Trustworthy security at the network edge demands fine-grained and temporally robust web attack detection to support autonomous threat response under the latency and communication constraints of distributed networked systems. Existing data-driven methods fail to deliver high-performance attack type classification as they either process HTTP fields independently or treat the entire request as a flat unstructured sequence, obscuring field-specific attack semantics. Moreover, the critical challenge of temporal distribution shift in dynamic network environments has received no systematic treatment in prior web attack detection work. To address both challenges, we propose DAMF-Net, a Domain-Adversarial Multi-Field Network that independently encodes HTTP fields via a shared SecureBERT encoder and fuses their representations through a field-aware attention mechanism that adaptively weights each field per sample, providing interpretable field-level evidence to support trustworthy autonomous threat response. An adversarial multi-source domain generalization module further promotes temporally invariant representations across multiple data periods. To support fine-grained evaluation and temporal robustness validation, we construct TMF-HAD, a 13-month real-world HTTP attack log dataset with fine-grained labels for seven attack categories from a production cloud Web Application Firewall (WAF) platform. Extensive experiments demonstrate that DAMF-Net achieves an average intra-domain accuracy of 98.19% and a macro-averaged F1 of 97.82% over 13 months, ranking first or co-first in 9 of 13 accuracy evaluations and 6 of 13 macro-F1 evaluations, while consistently outperforming the unadapted baseline under severe temporal drift. On three public benchmarks, DAMF-Net matches or surpasses prior state-of-the-art methods.
Edge cloud computing in the Industrial Internet of Things (IIoT) enables latency-sensitive tasks from IIoT terminals to be offloaded to distributed edge data centers (EDCs). This paper proposes an agentic artificial intelligence (AI)-assisted Stackelberg game framework to address the task offloading and resource allocation (TORA) problem. Our goal is to minimize the total processing delay of tasks while guaranteeing their delay thresholds. In this framework, LAN bandwidth and EDC computing capabilities are treated as commodities. Tasks act as non-cooperative Stackelberg followers, utilizing a Lagrange multiplier-based algorithm to achieve closed-form theoretically optimal bidding strategies under given commodity prices. Meanwhile, an agentic Stackelberg leader adopts a twin delayed deep deterministic policy gradient (TD3) algorithm to dynamically adjust these unit-prices for optimal system performance. Experimental results demonstrate that the proposed framework features extremely low computational complexity for online TORA problems, minimizes total processing delays, and significantly reduces the occurrence probability of events that at least one security-related task exceeds its delay threshold.
A channel knowledge map (CKM) provides location-specific channel priors and can reduce the overhead of real-time channel state information (CSI) acquisition for 6 G environment-aware communications. In practice, CKM generation is often constrained by sparse and noisy measurements due to the high cost of wireless data collection. In this paper, we propose PDiff, a physics-informed conditional diffusion framework for CKM generation under sparse observations. Specifically, PDiff incorporates an analytical free-space propagation prior to capture the dominant distance-dependent attenuation trend, and combines it with environmental geometry, observation masks, and sparse observations as structured conditions. These conditional inputs guide the generation process with explicit propagation-aware, environmental, and measurement constraints. To improve inference efficiency, we further develop Prop-Cache, a training-free acceleration mechanism that reuses slowly varying intermediate features across denoising steps to reduce redundant computation during sampling. Experiments on RadioMapSeer demonstrate that PDiff outperforms a wide range of baseline methods for CKM generation.
The radio map characterizes the spatial distribution of spectrum resources within a region of interest and plays an important role in wireless network planning and spectrum management. In practice, observations are often sparse, making accurate radio map reconstruction challenging. Although deep learning-based methods can recover a radio map from sparse samples, they often suffer from two major limitations: a strong dependence on large amounts of training data and reconstructed results that may deviate from the actual physical distribution. To address this issue, this paper introduces dictionary factors to characterize radio propagation properties at different spatial scales and formulates radio map reconstruction as a multi-scale dictionary factor learning problem. Based on this formulation, we propose RadioMSDL-Net, a radio multi-scale dictionary learning network. The network consists of multiple layers with identical structures and progressively refines the radio map estimate in an iterative manner. In each layer, the dictionary factor update module learns propagation characteristics at different spatial scales, while the multi-level reconstruction module integrates cross-scale features to improve both the global structure and local details of the radio map. Extensive experiments demonstrate that RadioMSDL-Net consistently outperforms existing methods in reconstruction accuracy, computational efficiency, and cross-environment generalization.
Agentic Artificial Intelligence (AI) enables task execution by coordinating calls to service agents, which makes it attractive for edge services under diverse user requests in edge networks. However, this workflow introduces new privacy risks. First, transmitted call messages may expose sensitive content through explicit fields and values. Second, the observable calling sequence over the network may reveal user intent through the selection and ordering of called service agents. Existing edge agentic AI methods mainly improve efficiency through multi-agent coordination, model compression, or cloud offloading, while privacy defenses are still limited and often focus on input filtering, interface masking, or action constraints. To address these issues, we propose a two-layer privacy-preserving agentic framework with a device controller agent and a service agent pool. The device agent integrates three privacy modules: (i) The disclosure control module marks each field value pair as “keep”, “hide”, or “drop” to limit plaintext exposure. (ii) The variational information bottleneck encoding module replaces “hide” values with protected semantic hints to preserve utility. (iii) The decoy insertion module adds safe decoy calls under a per-subtask budget to reduce trace leakage. Experiments across diverse tasks and settings show that the proposed framework reduces both content leakage and trace leakage while maintaining task utility.
