Short video streaming has become a dominant content format on mobile platforms. Due to frequent user swipes and time-varying network conditions, video chunks may arrive late, leading to playback stalls or unnecessary bandwidth waste. Although multi-path transmission has been explored to accelerate video delivery, existing approaches mainly focus on transport-layer resource scheduling and optimize path utilization in isolation, without coordinating with application-layer video chunk preloading decisions, which limits their effectiveness. In this paper, we propose JUMP, a client-side multi-path adaptive transmission framework for short video streaming that jointly optimizes video chunk preloading and multi-path scheduling. We formulate the video chunk transmission process as a sequential decision-making problem, which is effectively addressed using an Asynchronous Advantage Actor–Critic (A3C) algorithm to learn adaptive joint policies. Experimental results show that, compared with the baseline schemes, JUMP reduces playback stalling by 7.49%–37.58% in the low-bandwidth scenario, improves average bandwidth utilization by 8.33%–39.68% across three bandwidth settings, and achieves Quality of Experience (QoE) gains of 11.72%–36.16%.
The deployment of vision-language models (VLMs) for on-orbit remote sensing is traditionally hindered by the significant computational and energy demands of the models and the restricted hardware capabilities of individual low Earth orbit (LEO) satellites. This paper proposes OrbitServe, a novel energy-aware distributed on-orbit inference framework designed to orchestrate VLM workloads across a dense satellite-cluster formation. By leveraging a tensor-parallel approach tailored for simultaneous lightwave information and power transfer (SLIPT) enabled optical inter-satellite links (OISLs), Goodput@SLO is maximized by (i) offline profiling, and (ii) an epoch-based planner that jointly selects TP degree and effective concurrency under energy constraints. Trace-driven evaluation with an interactive remote-sensing VLM workload shows that OrbitServe significantly improves sustainable serving capacity across diverse orbital energy regimes, and that SLIPT enlarges the feasible operating region only when communication occupancy is non-negligible.
Intent-based networking enables automated network configuration and management through high-level natural language instructions. The core challenge lies in accurately parsing the diverse user requirements into executable formal specifications. Existing approaches either rely on formal languages for efficiency or employ Large Language Models (LLMs) for direct configuration generation, yet both suffer from limited usability and insufficient parsing accuracy. To address this, we propose a DeBERTa-based joint model that explicitly captures the directional and path semantics of key trigger words via a trigger-aware module and introduces a gated interaction mechanism to inject configuration-specification-level semantics into the entity recognition branch, thereby enhancing discriminative capability across semantic roles. We evaluated our model on a self-constructed dataset against various natural language understanding (NLU) baselines and LLMs. Experimental results show that our approach improves the entity recognition F1 score and overall sentence accuracy by approximately 3 percentage points over the best NLU baseline, while achieving significantly lower entity recognition error rates than both traditional NLU models and LLMs.
As ransomware techniques evolve, their threats have extended to critical infrastructures such as the increasingly digitalized smart grid. The smart grid’s highly interconnected communication and control architecture increases system complexity and broadens the attack surface, making it a prominent target for ransomware. Characterized by stealth, rapid propagation, and destructive capability, ransomware poses severe challenges to traditional defenses that struggle to respond promptly and effectively. To address these limitations, this paper proposes an early detection and active defense approach for ransomware in smart grids. The method deploys customized deceptive resources based on behavioral patterns at different attack stages to induce early exposure of abnormal behaviors for rapid identification. A multi-node collaborative defense mechanism allows independent nodes to securely share threat intelligence, enabling coordinated adjustments of defense strategies. Additionally, a permission-driven access control mechanism is incorporated to address the potential bypass issues associated with relying solely on deception-based techniques. The proposed method provides efficient and generalized protection across the ransomware lifecycle, enhancing the overall resilience of smart grid systems.
Speech Analytics Systems are increasingly deployed in customer service and financial applications, where audio is collected on edge devices for centralized recognition and auditing. This workflow introduces substantial privacy risks because speech conveys both sensitive content and biometric identifiers. However, existing protections are often difficult to deploy because they require developers to integrate heterogeneous Digital Signal Processing and deep learning components, manually tune parameters, and continuously balance evolving privacy policies against downstream utility. We introduce PrivSpeech, an adaptive framework that reduces this burden through string-based privacy queries. PrivSpeech compiles high-level objectives into executable pipelines by automatically selecting and composing operators for content masking and identity protection. To balance privacy and intelligibility, we propose Adaptive Semantic-Preserving Voice Anonymization (ASP-VA), which optimizes transformation parameters to minimize identity leakage while preserving downstream utility. Furthermore, PrivSpeech improves efficiency through shared-prefix execution, which merges overlapping computations across concurrent queries. Extensive evaluations demonstrate that PrivSpeech effectively secures sensitive information while maintaining high utility for speech analytics tasks compared to existing baselines.
Joint entity-relation extraction is essential for transforming unstructured cybersecurity text into structured knowledge that supports downstream security analytics. While Transformer-based encoders have shown strong effectiveness for information extraction, real-world security applications often require models with reduced latency and memory footprint. This paper presents an empirical evaluation of lightweight Transformer models for end-to-end joint extraction in the cybersecurity domain. We employ a unified multi-task framework that shares a Transformer encoder across entity recognition and relation classification, and optimizes both tasks jointly under a single objective. We conduct a controlled comparison of multiple compact Transformer variants using consistent training settings and evaluation criteria, and analyze the resulting accuracy-efficiency trade-offs. Our best-performing model, JointER-ModernBERT, achieves 0.9032 macro-F1 on entity recognition and 0.8549 macro-F1 on relation extraction. These findings provide practical guidance for selecting compact Transformer backbones for joint extraction in cybersecurity, and offer an evidence-based reference for deploying structured information extraction under resource-constrained settings.
The rapid growth of Internet of Things (IoT) applications has led to a diverse range of connectivity requirements, driving the evolution of cellular IoT technologies. While 5G NR was initially designed to support three distinct service families, enhanced Mobile Broad-Band (eMBB), Ultra-Reliable Low-Latency Communications (URLLC), and massive Machine-Type Communications (mMTC), many emerging IoT deployments fall between these extremes and require moderate data rates and latency with low device cost and long battery lifetime. To address this intermediate segment, 3GPP Release 17 introduced Reduced Capability (RedCap) devices, also known as NR-Light, which constrain key UE capabilities (e.g., bandwidth, antenna configuration, and modulation order) while preserving native NR operation and 5G core integration. This paper provides a practitioner-oriented review of RedCap within the broader cellular IoT ecosystem. We position RedCap among competing connectivity options, summarize the Rel-17 capability profile and standardized complexity-reduction and power-saving mechanisms, and discuss the resulting trade-offs in coverage and spectral efficiency. Finally, we review the evolution towards further reduced-capability profiles, including enhanced RedCap (eRedCap) in Release 18 and directions towards Release 19, to support connectivity selection for mid-tier cellular IoT deployments across diverse vertical sectors.
The efficient deployment of deep neural networks (DNNs) in resource-constrained edge environments necessitates intelligent model partitioning strategies that can leverage distributed computing resources while minimizing end-to-end inference latency. Existing approaches often suffer from limitations such as reliance on simplistic binary (edge-cloud) partitioning, static schemes that fail to adapt to dynamic network conditions, and high computational complexity that hinders real-time optimization. To address these challenges, this paper reformulates the multi-node DNN partitioning problem as a Facility Location Problem (FLP). This strategic transformation provides a theoretically grounded and computationally tractable foundation for joint optimization of partition points and DNN layer deployment. Building upon this formulation, we develop Feasibility Pump-based Neural Partitioning (FPNP), which can determine near-optimal partitioning solutions in near-real time, even for large-scale networks. Our approach explicitly accounts for both computation and communication costs under heterogeneous resource constraints. Extensive experimental evaluations demonstrate the superiority of FPNP.
The proliferation of Internet-of-things (IoT) devices, connected vehicles, and emerging 6G services is driving unprecedented demand for ubiquitous connectivity and low-latency computing. While standalone terrestrial networks struggle to meet these requirements, Integrated Terrestrial and Non-Terrestrial Networks (ITNTN) have emerged as a unifying architecture that spans ground, aerial, maritime, and satellite segments for ubiquitous and reliable connectivity. Mobile Edge Computing (MEC) is expected to play a central role in this ecosystem by bringing computation closer to end users and enabling real-time, computation-intensive applications. This survey provides a concise overview of MEC-enabled ITNTN, summarizing recent advances across UAV, HAPS, maritime, and satellite platforms, and examining how computing resources are distributed and coordinated across the ITNTN layers. Unlike existing surveys that treat these domains separately, this work highlights the benefits and design considerations of a fully integrated MEC–enabled ITNTN architecture. The paper concludes by outlining open issues and emerging research directions that will shape next-generation computing in ITNTN.
The expansion in size of IoT networks, especially ones that are based on battery-powered, constrained CPUs and storage hardware, necessitates optimization efforts in terms of minimizing radio transmissions, as radio activity is some of the most power-consuming tasks that such devices can perform. For a multi-hop IoT network, in-network content caching is a promising solution to relieve the network of retransmissions and reduce latency. However, since the amount of storage space in constrained IoT devices is expected to be small, the number of pieces of content that can be cached at a time can be limited, resulting in a low cache hit probability. We present a lightweight Linear Network Coding (LNC) approach aimed at low-power devices. Pieces of content are linearly combined and subsequently decoded by the receiver, thus increasing the probability of innetwork cache hits. We implement our solution on a 30-node hardware testbed and evaluate it across 600 experimental trials. Our results indicate that the integration of LNC with in-network caching increases the cache hit rate by 14.3% and reduces network-wide transmissions by 3.08% on average compared to uncoded caching. These early findings highlight the potential of network coding as a practical energy-saving mechanism for constrained, multi-hop IoT deployments.
Enhancing anomaly detection in Network Intrusion Detection Systems (NIDS) is critical for the security and reliability of 5G networks. With the introduction of network slicing, Internet-of-Things (IoT) integration, and low-latency services, modern systems are confronted with a more complex attack surface, challenging traditional detection mechanisms. Hence, effective anomaly detection enables real-time identification of suspicious activities, which is conventionally supported by machine learning (ML). However, ML struggles to detect threats in dynamic and heterogeneous environments. In this paper, we tackle this issue by leveraging quantum ML (QML). Specifically, we propose a novel QML-based NIDS supported by basic/Variational Autoencoders (VAEs), Principal Component Analysis (PCA), and dataset subsetting for features and dimensionality reduction. Through experiments, we demonstrate that our proposed QML framework effectively simplifies input data, optimizing the number of qubits and depth of quantum circuits required to address anomaly detection, resulting in faster processing, while maintaining a high cyberattack detection accuracy, compared to baselines such as Support Vector Machines (SVMs). Moreover, an impact analysis emphasizes the criticality of data distributions, data size, and class distributions on the QML-based NIDS performance, thus presenting novel guidelines for the efficient use of QML for anomaly detection.
The large-scale Low Earth Orbit (LEO) satellite network is expected to be an infrastructure network for the future Internet. As part of 3GPP’s vision for non-terrestrial network (NTN) integration in 5G and 6G, LEO satellite networks will merge with terrestrial networks to provide ubiquitous connectivity—from ground to space, and oceans to land. However, significant challenges remain, with routing being one of the most difficult due to the dynamic nature of LEO constellations caused by rapid satellite movement. The widespread deployment of inter-satellite links (ISLs) has further intensified the need for standardized routing solutions. There are many proposals for routing solutions from both academia and industry. This paper analyzes existing proposals, and progress in IETF. It focuses on proactive routing as the most promising candidate for future standard solution by IETF. To make the existing proactive routing protocol to be used in LEO satellite network, we propose that significant enhancements are needed in three key areas: network partitioning, convergence dependency and flooding mechanism.
Vehicular networks are essential for enabling smart transportation systems and improving road safety, traffic management, and overall connectivity. As these networks evolve to support increasingly complex applications, the demand for high computational capacity continues to rise. Tasks such as real-time data processing and surveillance require efficient and scalable solutions. To do so, uncrewed aerial vehicles (UAVs) offer a flexible approach to meet these computational needs by enabling on-the-fly task offloading. Their integration with connected and autonomous vehicles (CAVs) further expands their potential in intelligent transportation systems. In this context, we investigate computation offloading to UAVs in vehicular networks. Specifically, we aim to maximize the successful offloading rate of CAV tasks to a multi-UAV and energy-constrained aerial platform through the optimization of UAVs’ launch locations, flight directions, and CAV-UAV associations. Given the complexity of the formulated problem, we propose two low-complex metaheuristic-based approaches, namely the bat algorithm (BA) and particle swarm optimization (PSO)-based methods, to solve it. Moreover, we adapt the iterative exhaustive-linear programming (IE-LP) solution, developed in [1], to the multi-UAV scenario. Through extensive simulations, we show that IE-LP provides the best performance with low complexity for small systems (number of UAVs below 3), while BA and PSO-based approaches are preferred for their low complexity and high scalability.
Feature selection for network intrusion detection remains challenging due to high-dimensional feature spaces, strong dependencies, and cross-dataset distribution shifts. Existing metaheuristic methods, including the Equilibrium Optimizer (EO), often exhibit premature convergence and unstable search behavior because they lack a unified mechanism to coordinate exploration, exploitation, and boundary handling. This paper proposes a network temperature driven equilibrium optimization framework (EO2), in which network temperature acts as a global scheduler to regulate search scale and enable a smooth transition from global exploration to local exploitation. Built upon this framework, EO2 integrates distance-aware directional modulation and a constraint-consistent stabilization strategy to improve convergence robustness in high-dimensional spaces. Extensive experiments on the UNSW-NB15 and CICIDS2017 datasets, including cross-dataset evaluations, show that EO2 consistently outperforms EO and representative metaheuristic baselines in detection accuracy, feature sparsity, and convergence stability, demonstrating the effectiveness of network temperature as a global coordination mechanism for network intrusion detection.
Mobile LiDAR technology is increasingly employed to generate detailed 3D point clouds of electrical transmission infrastructure, where power lines appear as slender, suspended elements. However, these lines are frequently fragmented or absent in the data due to occlusions caused by vegetation, built structures, or unfavorable scanning geometries, posing significant challenges for automated reconstruction in complex environments. This survey offers a comprehensive review of existing methods designed to reconstruct occluded or incomplete power lines from LiDAR data, classifying them into four main categories: geometric fitting techniques (e.g., parabolic and catenary models), machine learning (ML) and deep learning (DL) approaches, graph-based strategies, and hybrid models that integrate physical priors with data-driven inference. Each method is examined with respect to its robustness to occlusion, computation efficiency, data requirements, and reconstruction accuracy. The paper also discusses key evaluation metrics. It concludes by outlining open challenges and future research directions, including generative modeling under severe occlusion, multimodal data fusion, and real-time reconstruction for operational deployment in power grid monitoring systems.
Cloud-native service function chaining (SFC) faces challenges in orchestration complexity, fault-tolerance cost, and hardware acceleration. This paper presents Kube-SFC with three key contributions: (1) a two-level decoupled orchestration that reduces control-plane rule-configuration operations from O(N2) to O(N) for coordinated chain attachment points, enabling rapid deployment within seconds; (2) a backup-free fault-resilience mechanism for policy-designated non-critical VNFs through policy-based VNF criticality classification, which restores connectivity within seconds via dynamic bypass without standby replicas; (3) KubeVirt vDPA integration, an open-source solution that combines live migration and hardware-level performance, improving throughput by 3.4×. Experimental results show that Kube-SFC significantly improves operational agility while preserving high-performance forwarding.
We present trust-driven adaptive sampling (TDAS), a lightweight and efficient path validation scheme designed for resource-constrained Internet of Things (IoT). The key innovation of TDAS is its trust-driven adaptive sampling mechanism, which strategically validates packet flows based on dynamic node trust and a random factor, moving beyond conventional per-packet or fixed-probability approaches. This allows TDAS to focus validation overhead where it is most needed—on untrusted or unstable paths. The scheme employs grid-based network partitioning for scalable management and uses lightweight cryptographic operations (truncated hashes and MACs) for proof generation and verification. A non-cryptographic hiding strategy effectively conceals sampling states from adversaries. Security analysis confirms TDAS’s resilience against path deviation attacks. Evaluation results demonstrate that TDAS significantly outperforms state-of-the-art benchmarks (Hummingbird and EPIC). In simulations of a 100-node IoT network, TDAS increases system throughput by up to 15% and reduces average path validation time by approximately 20%. These performance gains are also validated on a physical IoT testbed using DAYU800 & DAYU200 development kits running OpenHarmony OS.
Low-Earth-orbit (LEO) satellite constellations are emerging as a promising infrastructure for uplink-dominated Internet of Things (IoT) services. However, rapid satellite mobility leads to highly time-varying user–satellite associations, frequent handovers, and limited onboard processing capacity, which jointly complicate uplink scheduling design. Moreover, uplink tasks often span multiple time slots, resulting in strong cross-slot coupling due to task queue evolution and association switching costs. This paper investigates online uplink scheduling for multi-user LEO satellite IoT networks under time-varying visibility and switching costs. We formulate a cross-slot uplink scheduling model that captures dynamic association, onboard processing constraints, and task backlog evolution. To address the resulting complexity, we propose a Lyapunov-based primal–dual online scheduling algorithm that operates without future system information and guarantees queue stability. Simulation results demonstrate that the proposed approach effectively reduces uplink delay, improves satellite load balancing, and suppresses unnecessary handovers compared with representative baseline schemes.
The convergence of the Metaverse with immersive retail is accelerating the adoption of Virtual Try-On (VTO) systems, transforming e-commerce into an interactive experience. In this context, we propose a deployment-oriented optimization pipeline for real-time 3D hand reconstruction for ring VTO over 5G edge AI infrastructure. High-capacity transformer backbones provide the geometric fidelity required for stable accessory alignment, but their inference cost and footprint hinder scalable edge deployment. Starting from a trained baseline, our pipeline combines dependency-aware structured pruning, accuracy-retention fine-tuning, and hardware-aware acceleration. We evaluate the resulting deployable configurations and quantify the trade-offs between footprint, inference latency under single- and multi-user scenarios, and reconstruction accuracy.
Key Value Indicators (KVIs) provide a decision oriented view of a service by summarizing how operational performance translates into stakeholder value, risk, and outcomes. However, in many domains KVIs are difficult to compute in practice because they require selecting relevant KVI categories, defining measurable Key Performance Indicators (KPIs), collecting KPI values, and applying consistent calculation logic, all of which is typically performed manually and inconsistently from unstructured service documentation. This paper presents KPI2KVI, a tool that transforms a natural language service description into computed KVI estimates by orchestrating a deterministic multi agent workflow powered by Large Language Models (LLMs) that (i) elicits missing service context, (ii) extracts and finalizes relevant KVI categories from a taxonomy, (iii) generates service specific KPIs with units and descriptions, (iv) collects KPI values through an interactive dialogue and also supports intelligent estimation for KPI values that are unavailable, and (v) computes interval valued KVI outputs (minimum, exact, maximum) with traceable explanations for each KVI code. Simulations with representative service descriptions demonstrate that KPI2KVI consistently produces a complete end to end mapping from description to KVI intervals and provides transparent calculation narratives that support post hoc auditing and interactive advisory queries.
