Last updated: 2026-11-20 05:01 UTC
All documents
Number of pages: 179
| Author(s) | Title | Year | Publication | Keywords | ||
|---|---|---|---|---|---|---|
| Yuan Gao, Xuelong Wang, Zhenguo Dong, Yong Zhang, Yanyan Han | C2R: Causal-Corrected Residuals for Network Traffic Forecasting in Proactive Network Management | 2026 | Early Access | Modeling Forecasting Timing Learning (artificial intelligence) Fasteners Windows Foundation models Libraries Technology Telecommunication traffic Network Traffic Forecasting Time-Series Foundation Models Causal Discovery Residual Correction Network Management | In large-scale networks characterized by complex spatio-temporal couplings, accurate traffic forecasting serves as a cornerstone for proactive network management, capacity planning, and dynamic bandwidth allocation. Recent deep learning-based time-series foundation models have demonstrated remarkable capabilities in capturing long-term temporal dependencies for automated operations. However, a critical gap remains as these architectures often overlook network-specific topology in management scenarios and lack explicit modeling of cross-variable structural dependencies associated with traffic variations. These limitations can affect forecast accuracy during abrupt traffic changes. To address this challenge, we introduce C2R, a novel framework that augments the Chronos-Bolt foundation model with a TabPFN-based residual correction mechanism. Specifically, C2R constructs a regime-aware LPCMCI graph library by clustering historical traffic windows and learning lagged structural dependencies from the pooled observations within each regime, and then maps the selected structural candidates to residual errors to refine horizon-specific traffic forecasts. Experiments on core-network traffic datasets show that C2R achieves dataset-dependent incremental forecasting gains over the frozen backbone at the predefined evaluation origin while providing model-level structural attribution through the dependencies used by the residual-correction module. | 10.1109/TNSM.2026.3755281 |
| Shyam Kumar Shrestha, Shiva Raj Pokhrel, Jonathan Kua | Adapting Large Language Models for TCP Fairness over Wi-Fi | 2026 | Early Access | Large language models TCP Modeling Training Fluid flow Information rates Throughput Wireless fidelity Learning (artificial intelligence) Management Transmission Control Protocol Congestion Control Algorithms Fairness Wi-Fi | Designing Transmission Control Protocol (TCP) Congestion Control Algorithms (CCAs) for heterogeneous Wi-Fi networks remains a challenge due to rapidly varying delay, loss, and contention dynamics. These effects often lead to CCA incompatibility, flow unfairness, and starvation among competing TCP flows. Although Deep Reinforcement Learning (DRL) has shown promise in mitigating these challenges, its slow convergence, high training costs, and limited generalization hinder practical deployments. In this paper, we propose TCP-LLM, a learning-augmented transport framework that leverages Large Language Models (LLMs) for compatibility-aware CCA selection. TCP-LLM encodes multivariate TCP time-series into token embeddings and performs single-step discrete CCA decisions through a lightweight LLM decision head. Parameter-efficient low-rank adaptation (LoRA) reduces the number of trainable parameters by approximately two orders of magnitude. Evaluation using traces collected from a physicalWi-Fi testbed with competing TCP Cubic, Bottleneck Bandwidth and Round-trip propagation time (BBR), and Performance-oriented Congestion Control (PCC) flows shows that TCP-LLM achieves the highest mean, geometric-mean, and minimum per-flow throughput among the evaluated schemes. TCP-LLM also exhibits faster and more stable training convergence and supports low-latency bounded inference. These results demonstrate the feasibility of LLM-based sequence adaptation for compatibility-aware transport-layer supervision under heterogeneous wireless conditions. | 10.1109/TNSM.2026.3739364 |
| Soonbeom Kwon, Yusu Noh, Youngwoo Jang, Illyoung Choi, Byungchul Tak, In-geol Chun, Young-Kyoon Suh | Scalable and Robust Resource Provisioning via Adaptive Task Scheduling for Edge Devices | 2026 | Early Access | Schedules Scheduling Cloning Timing Educational institutions Computers Transcoding Videos Tail Edge computing Edge devices Edge server Resource augmentation Task distribution Kubernetes | Edge devices, such as wearables, drones, and CCTV systems, are vital for real-time data collection in urban intelligence. However, their limited computational and storage capacities pose significant challenges. While offloading to public clouds offers scalability, it often incurs high latency and operational costs. Conversely, centralizing workloads on edge servers may result in the underutilization of high-performance edge devices. To address these limitations, we introduce ERPF, a Kubernetes-based Edge Resource Provisioning Framework that augments the capabilities of heterogeneous edge environments. ERPF orchestrates dynamic volume provisioning, GPU-aware resource allocation, execution context migration, and adaptive task distribution to improve system flexibility and efficiency. Building on this, we propose a novel adaptive task scheduling technique, termed eATS, composed of three key mechanisms: (i) Partition Smoothing Scheme for stable task granularity control, (ii) Resilient Edge Reintegration for failure detection and task reassignment, and (iii) Competitive Task Cloning for speculative execution with fastest-result commitment. The proposed eATS scheme reduces task execution time by up to 27.6%, lowers partition size variability by 8.7×, and improves scheduling robustness across heterogeneous edge devices over the baseline. | 10.1109/TNSM.2026.3694238 |
| Xuejun Zhang, Hui Chen, Yexin Fan, Yanchun Hu, Jiami Guo, Xiaodong Yang | FedLoc-Guard: Cross-Round Behavior Modeling for Backdoor Defense in Federated RSSI Fingerprint Localization | 2026 | Early Access | Federated learning Modeling Location awareness Servers Received signal strength indicator Automatic speech recognition Fingerprint recognition Timing Educational institutions Cleaning Federated learning indoor localization Wi-Fi fingerprint localization backdoor attack cross-round behavior modeling robust aggregation | Federated learning provides a privacy-preserving paradigm for Wi-Fi RSSI fingerprint-based indoor localization, but remains vulnerable to stealthy backdoor attacks. In federated RSSI localization, environmental variation, device heterogeneity, and sampling-condition differences cause benign client updates to fluctuate across communication rounds, complicating malicious-update discrimination based on single-round statistics. To address this challenge, we propose FedLoc-Guard, a server-side defense framework that models client reliability from server-observable behavior across rounds. FedLoc-Guard estimates instantaneous reliability from directional consistency, relative update magnitude, and normalized turnaround-time consistency, updates client reliability across rounds, and uses the resulting reliability states for high-reliability client retention and risk-constrained aggregation with pre-normalization weight clipping. The method requires neither previously trusted clients nor an additional clean server-side reference set. Experiments on UJIIndoorLoc and Mall under IID and non-IID settings demonstrate that FedLoc-Guard substantially reduces attack success rates against BadNets, Scaling Attack, DBA, and an A3FL-inspired adaptive backdoor attack while maintaining useful localization performance under the evaluated settings, thereby improving the reliability and trustworthiness of federated indoor localization services. | 10.1109/TNSM.2026.3755818 |
| Edson José Pacheco, Mauro Sérgio Pereira Fonseca, Carlos Marcelo Pedroso | Interpretable Multi-Plane Intrusion Detection for 5G Networks with Adaptive Feature Engineering | 2026 | Early Access | Planing Modeling Recording Windows Modules (abstract algebra) 5G mobile communication Fluid flow Floods Timing Training Intrusion detection 5G security explainable machine learning cybersecurity feature engineering | Fifth-generation mobile networks expose multiple protocol planes to cyberattacks, yet intrusion detection research evaluates one feature representation, often on one plane, and relies on opaque classifiers that hinder root-cause analysis by network operators. We present an intrusion detection pipeline that applies one framework to every 5G network plane and trains an independent detector for each. The pipeline evaluates three feature modules (per-flow statistics, rolling temporal differentials, and windowed complexity descriptors) on each plane and selects one by a fixed rule under blocked and purged cross-validation, in which no training sample shares a record with a test sample. An Explainable Boosting Machine (EBM) classifies the traffic and produces auditable shape functions that reveal which traffic characteristics distinguish each attack class. Validation covers four public datasets, one of which contributes two planes, and spans user, mixed data, control, signaling, and service-based traffic. On these datasets, committing every plane to per-flow statistics, the module most common in the literature, costs 0.2493 F1 Macro on average and 0.6533 on NAS signaling. The EBM stays within 0.0058 F1 Macro of the best tree ensemble on every plane, and after Holm correction the cost of the additive restriction is not significant on any plane. On four planes, its mean F1 Macro is at or above that of six deep architectures. The service-based plane remains the hardest, at 0.5136 F1 Macro. These results indicate that interpretability and multi-plane coverage are achievable, and the pipeline accommodates further modules and planes as new attack surfaces emerge. | 10.1109/TNSM.2026.3755668 |
| Yichen Zhang, Di Chen, Kaijie Zhu, Zekun Tao, Wei Lin, Rui Zheng | Beyond Valley-Free: Transformer-based AS Path Inference Using Path Embedding | 2026 | Early Access | Border Gateway Protocol Routing Ranking (statistics) Modeling Training Transformers Measurement Conferences Heart rate Probability AS path inference inter-domain routing Border Gateway Protocol network traffic analysis | Gaining insight into autonomous system (AS)-level paths is essential for understanding Internet routing behaviors and enhancing inter-domain routing security. Due to increasingly complex Internet topology and limited measurement access (e.g., router-side ICMP restrictions), few AS paths can be accurately observed, making AS path inference a necessity. However, most existing inference methods are developed based on the valley-free constraint. This limits their ability to capture real-world dynamic routing behaviors. In this paper, we propose PathRank, a Transformer-based method that infers AS paths from a source-AS to a destination-prefix through probability-based ranking, thereby closely reflecting real-world inter-domain routing complexities. From observable BGP data, PathRank generates candidate paths, encodes them as embeddings for a self-attention mechanism to capture cross-hop dependencies, and employs a prefix-length-weighted loss function during training to reflect real-world traffic impacts.We conduct experiments on large-scale public BGP datasets covering IPv4 and IPv6. PathRank places the observed BGP path among the top-10 ranked candidates in 96.82% (IPv4) and 96.65% (IPv6) of test cases, outperforming the strongest external baseline by 4.20 and 11.72 percentage points, respectively. On non-valley-free cases, where valley-free-constrained methods degrade significantly, PathRank maintains inference performance. | 10.1109/TNSM.2026.3752819 |
| Ivette Kembely Carrera Manosalvas, Paulo Freitas De Araujo-Filho, Georges Kaddoum, Hadi Otrok | Unsupervised Hybrid Quantum-Classical Generative Adversarial Network-Based Intrusion Detection System for 6G-enabled Internet of Things | 2026 | Early Access | Modeling Signal detection Generative adversarial networks Timing Internet of Things Training Architecture Computer architecture Machine learning Fluid flow 6G Internet of Things (IoT) Intrusion Detection Systems (IDS) Generative Adversarial Networks (GAN) Quantum Machine Learning (QML) Security | The emergence of sixth-generation (6G) mobile networks contributes to the widespread adoption of critical Internet-of-Things (IoT) applications, such as telesurgery and autonomous vehicles. However, such advancements expand attack surfaces and make traditional security mechanisms insufficient. While machine learning (ML)-based intrusion detection systems (IDSs) have been increasingly adopted to detect attacks in such heterogeneous and dynamic environments, their complex architectures often result in high computational costs and latency, making them unsuitable for the real-time demands of 6G-enabled IoT applications. To address these challenges, we propose a novel unsupervised quantum generative adversarial network (QGAN)-based IDS. Our solution enhances the GAN generator’s ability to model complex data distributions by combining a parametrized quantum circuit with a classical neural network (NN), while maintaining a lightweight classical discriminator (fewer than 2.000 parameters) at the edge. This design significantly reduces computational complexity and minimizes detection times, since only the classical discriminator is used during inference. Experimental results demonstrate that our solution achieves competitive detection rates with an ultra-low inference latency of 0,22 ms, outperforming state-of-the-art GAN-based and transformer-based IDSs by over 2,81 times. Therefore, our architecture provides a resource-efficient security framework that delivers the ultra-low latency and high detection rates essential for next-generation IoT applications. | 10.1109/TNSM.2026.3747752 |
| Shengye Gong, Hui Yang, Qiuyan Yao, Xin Li, Tiankuo Yu, Wenxin Liu, Yang Zhao, Yucong Liu, Jie Zhang, Mohamed Cheriet | Resilient Hierarchical Polyhedron-based Configuration Protection for Large-Scale Multi-Domain Optical Networks | 2026 | Early Access | High performance computing Joining processes Protection Topology Costing Optical fiber networks Algorithms Modeling Architecture Computer architecture Network survivability multi-domain optical network network protection hierarchical polyhedron-based protection | Large-scale multi-domain optical networks are increasingly exposed to spatially correlated and concurrent failures, while monolithic pre-configured protection structures incur rapidly increasing global search complexity and protection-resource costs under heterogeneous physical constraints. To address this challenge, this paper proposes a resilient Hierarchical Polyhedron-based Configuration (HPC) architecture that decomposes a monolithic global survivability requirement into adaptive intra-domain connectivity constraints and a compact inter-domain connectivity constraint. To alleviate resource imbalance caused by topology-only network partitioning, a Topology-Resource Aware Domain Partitioning (TRADP) algorithm is developed to construct protection-supportable sub-domains by jointly considering topological structure and physical resource distribution. To avoid imposing a rigid protection level on heterogeneous domains, HPC assigns adaptive intra-domain connectivity targets according to local failure exposure and physical resource supportability. To reduce the complexity of cross-domain protection while preserving physical feasibility, a super-node abstraction and physical bridge-link mapping mechanism is employed to construct the inter-domain protection backbone under capacity, spectrum, and SRLG constraints. Extensive evaluations on networks of up to 1000 nodes and representative GEANT and NSFNET topologies demonstrate that HPC-TRADP maintains a recovery rate of 92.94% at 1000 nodes and 85.82% even when 30% of physical links fail concurrently. It further reduces protection overhead by up to 79.60% relative to p-KECS while maintaining efficient recovery paths, demonstrating a favorable balance between survivability, resource efficiency, and scalability. | 10.1109/TNSM.2026.3750417 |
| Rahman Abdul Rafi, Christian Stephens, Shakil Ahmed, Ibrahim Almazyad, Ashfaq Khokhar | Asynchronous SLM Orchestrator for Mission-Critical Tactile Communications | 2026 | Early Access | Modeling Costing Optimization Delays Timing Labeling Jitter Architecture Computer architecture Radio frequency 6G+ Tactile Internet Mission Critical Systems URLLC SLM Cross-Layer Optimization | Mission-critical tactile applications such as remote surgery require sub-10 millisecond (ms) one-way latency and deadline-miss rates below 0.1% to preserve haptic stability, yet existing cross-layer optimization methods struggle to adapt transport, redundancy, and application parameters under wireless conditions without violating time constraints. To address this gap, we present SLM-COP, a two-tier edge architecture that decouples real-time execution from adaptive policy selection. A fast dataplane executor applies pre-validated cross-layer presets at 1–10 ms intervals while an asynchronous small language model (SLM) meta-controller runs every 50–200 ms to interpret telemetry and language-encoded objectives to propose configurations from 72 validated presets spanning FEC level, congestion control, frame aggregation, sampling rate, and packet-loss concealment. Using trace-driven experiments with da Vinci surgical kinematic traces from the JIGSAWS dataset across 5,000 emulated network conditions, the SLM meta-controller achieved mean cost 8.93 under Top-1 selection on 1,000 held-out scenarios, improving over rule-based heuristics by 53% while maintaining 100% feasibility with DMR ≤ 0.1%. Further, oracle analysis shows 85% of scenarios admit multiple near-optimal presets within a 10% cost margin. This facilitates Top-K shortlist evaluation, under which SLM-COP reaches cost 7.08 at Top-5 versus 6.74 for oracle. Batched inference over all 72 presets requires 100 ms, aligning with the slow-loop budget and preserving the fast-loop real-time guarantees. Thus, we demonstrate that asynchronous SLM-based orchestration can provide adaptive cross-layer control for deploy-able tactile internet systems over emerging 6G+ infrastructure. | 10.1109/TNSM.2026.3755916 |
| Xiaolong Wang, Haipeng Yao, Wei Zhang, Tianle Mai, Hao Hao | Shared-Link-Aware Redundant Routing and CQF Flow Admission in Time-Sensitive Networks | 2026 | Early Access | Fluid flow Joining processes Schedules Timing Scheduling Optimization Modeling Routing Decoding Maintenance engineering Time-sensitive networking frame replication and elimination for reliability shared-link reliability redundant routing flow admission cyclic queuing and forwarding | Time-sensitive networking (TSN) requires deterministic latency and reliable delivery for industrial traffic. Frame replication and elimination for reliability (FRER) improves fault tolerance, but sparse topologies may force redundant routes to share links. The resulting overlap couples path failures with replicated cyclic queuing and forwarding (CQF) demand, making routing-layer reliability insufficient to ensure feasible admission and cycle allocation. We formulate joint redundant routing, weighted flow admission, and CQF resource allocation as a lexicographic optimization that maximizes admitted weight before minimizing resource use. Exact inclusion–exclusion evaluates shared-link reliability. Heuristic Optimization for Reliable Routing and Admission (HORRA) combines reliability- and deadline-feasible path-set libraries, a resource-aware CQF decoder, admission-gain-guided reordering, and bottleneck-guided destroy-and-repair. With a budget of 310 decoder evaluations, HORRA improves weighted admission by 1.7–3.3% relative to the one-shot baselines. Compared with the particle-swarm baseline, it reduces search time by 64.2% and 60.4% under moderate and severe contention, respectively. For identical admitted-flow sets, it reduces the resource objective by 11.65–16.84% relative to the two iterative baselines. The results demonstrate the benefit of coordinating shared-link reliability with CQF resource allocation. | 10.1109/TNSM.2026.3755712 |
| Hongyuan Cheng, Lihua Yin, Tianqing Zhu, Hongyu Yang, Dexin Zhu, Weixiang Jiang, Nan Wei | Pattern2Bot: A malicious server IP detection model based on self-trained heterogeneous graph neural network | 2026 | Early Access | Labeling Modeling Signal detection Training Botnet Servers IP networks Sequential analysis Conferences Security Botnet detection Self-supervised Graph neural network Command & Control service | Botnets play a key role in network attacks such as Distributed Denial-of-Service (DDoS) attacks, spamming and click fraud. Command & Control server is an important node for botmaster to control bots and distribute attack commands. Bot disguise as normal domain requests through DNS traffic, but leave malicious records in DNS logs containing client IP, query domain, server IP and other complex interactions. This information provides important clues for detecting those botnets. In the current study, researchers detect botnets through multiple perspectives, including abstracting DNS communication records into graphs and using graph neural network models for detection. Conventional approaches often necessitate massive labeled datasets for model training, which substantially incurs training costs. Moreover, despite enriching node embeddings via topological relations, graph models often suffer performance degradation when confronted with structural semantic obfuscation from evasion techniques like double-flux. In order to understand the abstract and complex relationship patterns in DNS records while reducing label usage to lower training costs, we propose a botnet Command & Control server detection model based on DNS traffic. This is a self-training heterogeneous graph neural network model that does not require a large number of labels. By integrating the high-order relational semantics derived from multiple meta-paths with the local topologies of target nodes, the model achieves finer-grained structural awareness and semantic abstraction, thereby enhancing detection performance. We validate the effectiveness of the model in a real dataset, and the results show that the model has a better understanding of the special structure possessed by the IPs of Command & Control servers. | 10.1109/TNSM.2026.3741111 |
| Pengbo Li, Xingwei Wang, Hao Lu, Qiang He, Donghong Han, Saru Kumari, Keqin Li, Sajal K. Das | FedC3P: Compatibility Cost Sharing and Congestion-Aware Posted Pricing for Heterogeneous Multi-Cloud Federations | 2026 | Early Access | Resource management Clouds Costing Pricing Investment Optimization Modeling Routing Delays Virtual machines Cloud federation compatibility investment congestion-aware posted pricing cost sharing provider participation resource allocation | Heterogeneous multi-cloud federations pool resources across providers. Compatibility investment reduces service costs caused by incompatibility, but returns depend on workload allocation under capacity and congestion. Existing allocation and pricing studies generally fix compatibility, leaving investment and cost sharing undetermined. We propose a mechanism for compatibility cost sharing and congestion-aware posted pricing in heterogeneous multi-cloud federations (FedC3P). FedC3P updates compatibility investment across providers on the slower time scale and allocates elastic workloads on the faster time scale. For fixed investment, FedC3P maximizes federation welfare subject to demand and regional capacity constraints and derives resource prices from optimality conditions. Given these prices and service costs from incomplete compatibility, independent service quantity and routing decisions by workload classes implement the welfare-maximizing allocation. FedC3P then shares compatibility costs, checks the feasibility of provider participation under budget balance, and determines provider payoffs through weighted Nash bargaining when distributable surplus is positive. We establish an O(T−1/2) saddle-gap bound for inner averages over T iterations and first-order stationarity of every outer accumulation point under sufficient conditions. At the default setting, semi-synthetic experiments on 277 cloud-region records show welfare gains of 18.9% and 30.8% over Routing-NUM and TD3-Routing; median runtime remains below 5 s at the largest scales. | 10.1109/TNSM.2026.3754272 |
| Wuzhao Li, Weian Guo | Switch-Coupled Scheduling for Energy-Bounded Heterogeneous Edge AI on Commodity Ethernet | 2026 | Early Access | Energy Modeling Joining processes Timing Poles and zeros Dispatching Windows Loading Scheduling Management edge AI energy-constrained scheduling Ethernet switching Lyapunov optimization serverless computing multi-tenancy | Heterogeneous edge-AI clusters behind commodity Ethernet must jointly manage compute queues, shared links, runtime residency, and energy budgets. We formulate request admission and assignment with per-hop FCFS resources, byte-limited TTL–LRU caches, temperature-dependent service, power uncertainty, deadlines, and tenant shares. Switch-Coupled Dual Scheduling (SCD) combines deadline feasibility, cost-aware energy credits, and device plus tenant virtual debts. Analysis establishes restricted-problem hardness, a state-correct finite-horizon certificate, and guarded commitment bounds. Nominal trace-driven simulations show higher offered on-time success and lower mean response than compute-first. These gains accompany higher task energy, admitted tail latency, and mean peak age of information. The success gain over matched drift-plus-penalty scheduling is incremental. Shared link state limits queue escalation under severe simulated bottlenecks. Elastic tenant allocation restores service under tight budgets while retaining device commitment bounds. Repeated hardware experiments compare scheduling under common resource constraints, execute trained neural and tree models on heterogeneous boards, and measure Ethernet path contention. The results support competitive deadline-oriented control with explicit service and energy trade-offs. | 10.1109/TNSM.2026.3752827 |
| Vincent Sheh, Yao Chiang, Hung-Yu Wei | Two-Timescale Service Orchestration for Dependable Multi-Region Edge Analytics under Adversarial Workloads | 2026 | Early Access | Modeling Delays Routing Costing Inspection Central Processing Unit Receivers Signal detection Degradation Accuracy Edge computing resource management task offloading intrusion detection systems Yo-Yo attacks multi-agent reinforcement learning | Adversarial workloads can saturate the shared CPU budget of edge nodes, undermining the continuous availability that safety-critical analytics require. Co-locating an inline intrusion detection system (IDS) can filter malicious traffic but creates contention with analytics inference over the same budget, producing a defense-service resource tradeoff. This tradeoff is destabilized by two mechanisms: adaptive Yo-Yo attacks that exploit container scaling delay, and load-based offloading that ignores cross-region IDS mismatch. We propose a two-timescale orchestration framework that jointly minimizes service-level objective (SLO) violations, residual malicious leakage, and benign quality degradation. A long-timescale multi-agent reinforcement learning policy partitions CPU between inspection and inference, internalizing scaling delay and coordinating across regions to suppress oscillatory reallocation. A short-timescale heuristic routes inspection traffic to receivers with compatible detection models, preventing the mismatched-IDS errors that load-based routing would induce. Compared with a representative state-of-the-art baseline, the framework achieves up to 68.8% lower SLO violation rate, 90.9% lower residual malicious leakage, and 64.6% fewer reallocations. | 10.1109/TNSM.2026.3753027 |
| Zhaogang Shu, Dawei Lin, Haihui Xie, Shuwu Chen | PeriSTNet: A Topology-Anchored Residual Graph Evolution Model for Multi-Step Traffic Matrix Forecasting | 2026 | Early Access | Forecasting Modeling Matrices Learning (artificial intelligence) Timing Topology Educational institutions Conferences Sequences Sequential analysis Traffic Matrix Forecasting Spatio-Temporal Graph Neural Networks Residual Graph Evolution Topology Anchoring Wide Area Networks | Accurate multi-step Traffic Matrix (TM) forecasting enables Wide Area Network (WAN) operators to plan capacity, anticipate congestion, and assess traffic-engineering actions before service quality is affected. Existing spatio-temporal graph neural networks often rely on a fixed physical topology, whereas fully data-driven and long-sequence models may not explicitly retain network-structure priors. These choices can miss demand correlations induced by routing and application behavior, while recursive decoding can accumulate error at longer horizons. We propose PeriSTNet, a topology-anchored residual graph evolution framework that learns a sparse, phase-conditioned predictive dependency graph around a physical topology prior. This graph is integrated into a Dual-Graph Fusion Gated Recurrent Unit, and a Spatio-Temporal Shortcut stabilizes recursive rollouts. Across Abilene, GÉANT, and CERNET, PeriSTNet achieves lower aggregate MAE, MAPE, and RMSE than the evaluated graph-based and Transformer baselines; horizon-wise, multi-seed, and efficiency analyses support this result under one reproducible protocol. By coupling dynamic predictive dependencies with a stable physical backbone, PeriSTNet provides a lightweight forecasting component for proactive capacity planning and traffic-engineering workflows. | 10.1109/TNSM.2026.3753266 |
| Huiying Jin, Jianyong Zhang, Jiajia Li, Hai Dong, Pengcheng Zhang, Jian Zhou | Dynamic Resource Allocation for Multi-Type Tasks in Mobile Edge Computing | 2026 | Early Access | Resource management Optimization Algorithms Timing Internet of Things Modeling Service level agreements Collaboration Elementary particles Bismuth Multi-type tasks Accurate matching scheme Multi-Node collaboration Dynamic resource allocation | We propose MD-SFLAPSO (/ˌemdi ˈsflæpsəʊ/), a Multi-type task-oriented Dynamic resource allocation method based on Shuffled Frog Leaping Algorithm and Particle Swarm Optimization for the mobile edge computing environment. MD-SFLAPSO offers an optimized resource allocation strategy. It incorporates a precise matching scheme that accounts for task requirements, collaborative node capabilities, spatial location correlation, latency constraint satisfaction, and resource supply-demand matching degree, to achieve effective collaboration. Additionally, we introduce a multi-node collaborative dynamic resource allocation strategy by leveraging user mobility patterns. Our approach is validated through a series of experiments using public data sets. The results show that MD-SFLAPSO reduces the average SLA violation rate by 8.8 percentage points from the pre-optimization level, improves average resource utilization by 9.3 percentage points over SFLA-RA, and adapts well to different resource configuration scenarios. | 10.1109/TNSM.2026.3751416 |
| Mario Chahoud, Hani Sami, Ahmad Hammoud, Rabeb Mizouni, Hadi Otrok, Jamal Bentahar, Azzam Mourad, Chamseddine Talhi | Multi-agent Deep Reinforcement Learning Scheme for On-Demand Federated Learning | 2026 | Early Access | Modeling Federated learning Training Learning (artificial intelligence) Convergence Accuracy Internet of Things Resource management Servers Optimization Federated Learning Multi-Agent Deep Reinforcement Learning On-Demand Federated Learning Client Selection Mobility-Aware Learning Partially Observable Markov Game Centralized Training and Decentralized Execution Edge Computing | The rapid growth of intelligent IoT devices and mobile edge infrastructures has generated massive decentralized data, necessitating efficient and privacy-preserving learning. Federated Learning (FL) enables distributed model training without sharing raw data; however, its performance degrades in dynamic mobile environments due to unpredictable availability, stragglers, and frequent dropouts. In On-Demand FL, training services and models are deployed only when needed rather than relying on always-active infrastructure, requiring orchestrators to adapt to client arrivals, departures, and mobility while efficiently allocating limited resources and maintaining training continuity. Existing client-selection methods rely on static heuristics or centralized Reinforcement Learning models that struggle to scale and adapt to such dynamics. To address these challenges, we propose a cooperative Multi-Agent Deep Reinforcement Learning (MADRL) framework for client selection in On-Demand FL. Multiple orchestrators act as agents under a Centralized Training and Decentralized Execution (CTDE) paradigm. The problem is formulated as a Partially Observable Markov Game (POMG), capturing mobility, partial observability, and distributed decision-making. A mobility-aware reward integrates accuracy, resource efficiency, and a Useful Handoff Bonus to preserve learning continuity. Containerization further enhances adaptability. Experiments demonstrate faster convergence, improved resource utilization, and enhanced robustness under real mobility traces. | 10.1109/TNSM.2026.3750310 |
| Mohammad Saberi, Ali Ahmadi Esfidi, Mahdi Dolati, Naeimeh Omidvar | FlexINA: In-Network Aggregation for Accelerating Distributed Machine Learning with Flexible Routing | 2026 | Early Access | Modeling Switches Timing Routing Machine learning Training Runtime Memory Loading Topology Distributed Machine Learning Programmable Networks In-Network Computing | The growing scale and complexity of modern machine learning (ML) models have rendered distributed training indispensable. However, as model dimensions expand, communication overhead increasingly dominates the overall training time, creating a critical performance bottleneck. While techniques such as gradient compression offer partial relief, they often incur substantial computational overhead or require specialized hardware. In-network aggregation (INA) has emerged as a compelling alternative, offloading partial computations—such as gradient reduction—directly onto programmable network switches. Yet, current INA frameworks predominantly rely on rigid, static routing schemes or uncoordinated, opportunistic aggregation, fundamentally limiting their capacity to exploit the full potential of programmable data planes. To address this limitation, we formulate the Joint Scheduling-Routing for In-Network Aggregation (JINA) problem. JINA introduces a time-slotted execution model to enable fine-grained temporal scheduling synchronously with spatial routing. By leveraging source routing, it dictates perpacket forwarding paths while explicitly coordinating multi-hop partial aggregation across the network. Furthermore, we propose FlexINA, a scalable optimization framework that solves the JINA problem through a three-phase dimensionality reduction strategy: host-level pre-aggregation, topological switch filtration, and temporal window partitioning. Extensive evaluations across fifteen diverse network topologies demonstrate that FlexINA reduces the total network load by up to 8.3% compared to INA baselines, while maintaining high computational efficiency. | 10.1109/TNSM.2026.3740503 |
| Franck Messaoudi, Luhan Wang, Abdelkader Mekrache, Adlen Ksentini, Bingxuan Li, Jialei Su, Sofiane Messaoudi, Salim El Ghalbzouri | The Brewing Storm in 5G’s Data Plane: Design and Evaluation of a High-Performance eBPF/XDP-Based User Plane Function | 2026 | Early Access | Quality of service Fluid flow Kernel Information rates Throughput Planing 5G mobile communication Linux Filtering Filters 5 th Generation Mobile Networks (5G) User Plane Function (UPF) QoS Enforcement Rule (QER) Quality of Service (QoS) extended Berkeley Packet Filter (eBPF) eXpress Data Path (XDP) Traffic Control (tc) Queuing Discipline (qdisc) | This paper presents the design and implementation of a novel 5G UPF leveraging eBPF technology to meet the stringent performance and programmability requirements of emerging 6G systems. Traditional UPF implementations often struggle to balance performance, flexibility, and resource efficiency-challenges particularly critical in CPU- and I/O-constrained edge environments. The proposed eBPF-based UPF architecture mitigates these limitations by embedding core functionalities, such as packet classification, forwarding, and QoS enforcement, directly within the Linux kernel via eBPF programs attached through XDP and tc hook points. Performance evaluation using TRex demonstrates that the proposed solution achieves competitive throughput, low packet loss, and efficient CPU utilization across traffic profiles. Moreover, it maintains full compliance with 5G Core Network standards. Comparative analysis with well-established open-source UPF implementations further underscores its advantages. This work highlights the potential of eBPF as a foundational technology for building next-generation, programmable UPFs optimized for edge cloud deployments in the 6G era. | 10.1109/TNSM.2026.3720812 |
| Shuang Zheng, Xing Zhang, Michael Sheng, Haixu Wang, Wenbo Wang | Beam Hopping Low Earth Orbit Satellite Resource Allocation for Differentiated Services and Robustness Analysis under Model Attacks | 2026 | Early Access | Beams Satellites Resource management Modeling Optimization Schedules Scheduling Low earth orbit satellites Algorithms Bridges LEO satellite communications deep reinforcement learning digital twin resource allocation adversarial attack | Beam hopping (BH)-enabled Low Earth Orbit (LEO) satellites play a pivotal role in next-generation communication networks, providing global coverage, improving spectrum efficiency, and supporting flexible adaptation to heterogeneous service demands. To fully exploit these capabilities, artificial intelligence (AI) techniques are increasingly employed for dynamic resource allocation and power management. However, limited onboard resources and potential adversarial perturbations pose challenges to both efficiency and robustness. To address these issues, we leverage digital twin technology to accurately capture the spatio-temporal dynamics of user–satellite visibility, providing precise state information for decision-making. Building on this, we formulate a joint optimization framework for BH scheduling and power allocation as a Markov Decision Process and propose the BRIDGE—BH with Reinforcement learning incorporating Integrated Dirichlet and Gumbel-TopK Exploration—which integrates a quality of service (QoS)-driven subchannel scheduling mechanism to ensure efficient and differentiated resource allocation. The model’s robustness is systematically evaluated under three classical adversarial attacks. Simulation results demonstrate that our approach achieves superior energy efficiency, service throughput, and fairness, while the robustness analysis shows stable performance under the considered bounded adversarial perturbations. | 10.1109/TNSM.2026.3710750 |