TNSM Searchable Index

Author(s)	Title	Year	Publication	Keywords
Ahlem Menaceur, Hamza Drid, Mohamed Rahouti, Yasmine Medjadba, Abdellah Chehri	Priority-Aware Dual-Queue Buffering for Reactive SDN Fault Management with Reduced Loss	2026	Early Access	Software defined networking Joining processes Loading	Software-Defined Networking (SDN) provides controller-driven flexibility for network and service management, yet reactive failure recovery often suffers from a recovery gap between fault detection and the installation of updated forwarding rules. Packets arriving during this interval are vulnerable to loss, while rerouted flows may overload specific links, undermining service continuity. To address these challenges, this paper proposes a dual-priority, queue-assisted reactive recovery mechanism that integrates traffic classification and buffering at the data plane. The design assigns flows to a high-priority queue (HPQ) or a low-priority queue (LPQ), ensuring that critical traffic is preserved while the controller completes rerouting. Evaluation on a 9-switch custom topology and the 16-node NSFNET backbone demonstrates that the proposed approach reduces packet loss (from 10.3% to 5.4% in the 9-switch topology and to below 1.9% in NSFNET) and improves post-failure load distribution, decreasing congested and surcharged links by 37.6% and 20.6%, respectively, compared to conventional reactive recovery. These results highlight that priority-aware buffering during the recovery gap enhances resilience, scalability, and operational efficiency for SDN fault and service management.	10.1109/TNSM.2026.3744288
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	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	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
Byungchang Kim, Shinyoung Lue, Seonyeong Hwang, Youhee Park, Yeji Lee, Jaemin Jeong, Peter Bloodsworth, Euiseok Hwang	GRIDS: Generalized Representation-based Intrusion Detection System via Federated Contrastive Learning	2026	Early Access	Modeling Labeling Intrusion detection	Existing AI-based intrusion detection systems are typically trained for specific network environments, leading to significant performance degradation when deployed in different settings or exposed to evolving attack strategies. We propose GRIDS (Generalized Representation-based Intrusion Detection System), a federated contrastive learning framework for cross-environment intrusion detection that keeps raw traffic data local. GRIDS employs a Transformer-based encoder trained with supervised contrastive learning across distributed network environments, learning a shared representation without exchanging sensitive traffic data. The encoder embeds network flow features into a latent space where semantically similar traffic patterns cluster together, enabling effective anomaly detection via k-nearest neighbor scoring. Evaluated on four heterogeneous NetFlow datasets, GRIDS outperforms existing federated IDS baselines on average with a single unified model. In leave-one-out cross-validation, a substantial portion of the detection capability of GRIDS transfers to network environments unseen during training, and the level of transfer varies across target networks. Deployment at such a network requires no labeled attack data or fine-tuning. It uses the benign traffic of the network as the kNN reference set and its unlabeled traffic for normalization and threshold calibration. The framework targets deployments in which a single model protects multiple heterogeneous networks.	10.1109/TNSM.2026.3748922
Jian Ye, Gaolei Fei, Jiajie Li, Xuemeng Zhai, Guangmin Hu, Ming Liang	Network topology inference via time-frequency consistency analysis of end-to-end traffic time series	2026	Early Access	Timing Topology Measurement	In this paper, we study the classical network tomography problem of tree topology inference using end-to-end measurement. Traditional methods rely on time series correlations to measure shared path lengths. However, time-varying and non-stationary background traffic often biases these measurements. To overcome this, we comprehensively explore end-to-end traffic time series via time-frequency consistency analysis. Specifically, we jointly characterize time series in both time and frequency domains, imposing a consistency constraint. This ensures embeddings of the same time series across domains remain close in the latent space, providing a robust description of traffic dynamics. Our contrastive learning approach effectively highlights similarities among flows on common paths and distinguishes those on disjoint paths, thereby mitigating background traffic interference. Extensive evaluations across widely used probing strategies and metrics demonstrate that our embedding significantly enhances time series correlation and accurately measures shared paths.	10.1109/TNSM.2026.3749558
Jian Liu, Jinhong Chen, Eldar Sabanovic, Jinliang Liu, Engang Tian, Oh-Min Kwon	Predictor-Enhanced Constrained Reinforcement Learning for Networked CAV Platoons: A Dual-Redundancy Scheme	2026	Early Access	Timing Vehicles Learning (artificial intelligence)	This paper investigates resilient tracking control for connected and automated vehicle (CAV) platoons under false data injection (FDI) attacks and prescribed spacing constraints. A kinematics-based state predictor is combined with a dynamic event-triggered mechanism (ETM) to provide smooth trajectories from sparse, FDI-corrupted measurements while reducing communication use. Prescribed performance control (PPC) maps the predicted spacing error into an unconstrained state for continuous-time actor-critic reinforcement learning (RL). The actor-critic controller learns an approximate tracking policy without identifying the unknown nonlinear drift. A locally sensed, saturated proportional-derivative fallback controller can override the learned command when the measured spacing error approaches the prescribed boundary. These two control paths form the dual-redundancy scheme. The analysis excludes finite-time accumulation of triggering instants, establishes input-to-state stability of the prediction-error subsystem, and gives conditional ultimate bounds for the prediction error and sampled adaptive weights under the stated input and approximation-domain conditions. Matched ablations and multi-seed tests evaluate communication use, fallback activation, and the responses to scalar and vector attacks.	10.1109/TNSM.2026.3748969
Qu Wang, Hao Wu, Xin Luo	A Highly Accurate Approach to Dynamic Network Representation Learning Based on Biased Nonnegative Tensor Ring Factorization	2026	Early Access	Modeling Tensors Ranking (statistics)	A large-scale dynamic network can be represented as a high-dimensional and incomplete (HDI) tensor, whose numerous entries are missing because it is impossible to acquire all interactions among massive numbers of nodes. Tensor network-based models have emerged as powerful tools for HDI tensor representation learning. Tensor ring (TR) decomposition represents a high-order tensor as a cyclic interconnection of third-order tensors, providing a compact tensor network structure for latent representation learning. However, conventional TR decomposition is designed for complete tensors and thus does not account for the substantial missing data in HDI dynamic networks, while existing TR-based representation learning models also have limited ability to model multidimensional link fluctuations. To address these issues, this work proposes a biased nonnegative tensor ring factorization (BNTRF) model with three main components: a) adopting a data-density-oriented modeling strategy that focuses computation on observed entries, thereby reducing the modeling cost of HDI tensors; b) designing a linear bias mechanism to capture multidimensional link fluctuations, thereby improving model stability and representation accuracy; and c) developing a self-adaptive parameter learning scheme to improve computational efficiency. Experimental studies on six real dynamic networks demonstrate that BNTRF outperforms several state-of-the-art models in learning accuracy while incurring lower computational costs.	10.1109/TNSM.2026.3748838
Martha Cash, Charlotte Fowler, Alexander M. Wyglinski	On the Role of Time Series Clustering in Traffic Matrix Prediction	2026	Early Access	Fluid flow Timing Modeling	This paper analyzes the role of time-series clustering in traffic matrix (TM) prediction. Traffic flows within a TM often exhibit heterogeneous behavior, which can reduce the effectiveness of global forecasting models that predict all flows jointly. To address this, we propose a clustering-based prediction framework that groups flows into smaller subsets and trains separate predictors for each group. We evaluate four traffic-flow representations for clustering, namely histogram, autocorrelation function (ACF), power spectral density (PSD), and naïve partitioning, and analyze how both the representation choice and the number of clusters affect prediction performance. Experiments using the publicly available Abilene and GÉANT datasets show that clustering consistently improves over global forecasting baselines, while remaining substantially less costly than local prediction. The results further show that most of the performance gain is achieved at moderate values of K, with diminishing returns as the number of clusters increases. Although different clustering representations produce different partitions of the traffic flows, they often achieve similar root mean squared error (RMSE). This suggests that the main benefit of clustering lies in decomposing the TM prediction task into smaller subproblems, while the exact cluster structure plays a more limited role in determining overall prediction accuracy.	10.1109/TNSM.2026.3747898
Mona Alghamdi, Atm S. Alam, Asma Cherif	A Context-Aware Action Filter for Deep Reinforcement Learning-Based Task Orchestration in Mobile Edge Computing	2026	Early Access	Filtering Servers Filters	The growing complexity and highly dynamic characteristics of Mobile Edge Computing (MEC) environments introduce considerable challenges for Deep Reinforcement Learning (DRL)-based task orchestration. With an increasing number of potential execution nodes, the set of available actions becomes substantially larger, leading to greater exploration complexity and exposing the learning agent to many infeasible or contextually inappropriate decisions. This paper proposes a Context-Aware Action Filter (CAAF) for Double Deep Q-Network (DDQN)-based task orchestration in MEC. CAAF dynamically evaluates candidate edge servers according to task characteristics, resource availability, network conditions, and system context. It combines a composite suitability assessment with an adaptive distribution-aware filtering mechanism that removes less-suitable actions when meaningful differences exist among candidates, while preserving action diversity when their suitability is comparable. The resulting reduced action space is then provided to the DDQN agent for task assignment. Simulation results demonstrate that CAAF-DDQN outperforms conventional DDQN, particularly under high-load conditions, reducing service time and task failure rate by approximately 20.0% and 45.7%, respectively. Scalability experiments with a larger number of edge servers further confirm its effectiveness under enlarged action spaces, achieving reductions of approximately 31.8% in service time and 89.5% in task failure rate without modifying the DDQN architecture or learning hyperparameters.	10.1109/TNSM.2026.3749201
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	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
Shubhani Aggarwal, Georges Kaddoum	QSB-SHA: Quantum-secured Blockchain for Smart Home Authentication	2026	Early Access	Authentication Smart homes Internet of Things	Today, IoT enables smart home users to remotely connect with household appliances via the Internet. Sensor-equipped smart devices are widely deployed in smart homes to collect and manage information regardless of user location. However, ensuring the secure and private accessibility of these devices in a centralized manner remains a major concern. The emergence of blockchain and quantum computing, with their enhanced information processing capabilities on decentralized platforms, offers a viable solution to this issue. Consequently, there is a need for cryptographic primitives that can withstand both quantum and classical attacks. To address this concern in the present study, we propose a quantum-secured blockchain (QSB) framework for smart home authentication (SHA). The proposed framework employs quantum key distribution (QKD) and maximally entangled states on a decentralized platform to enable secure key distribution and mutual authentication between smart home users and devices. Furthermore, our approach integrates quantum passwords and quantum digital signatures to verify and validate legitimate users, thereby ensuring security and privacy. The framework is simulated using the Automated Validation of Internet Security Protocols and Applications (AVISPA) tool for formal security analysis. We also conduct an informal security analysis and compare the performance of our proposed scheme with that of the previously proposed solutions. The results demonstrate that our framework outperforms existing approaches across various performance evaluation metrics.	10.1109/TNSM.2026.3747090
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	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
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	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
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	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
Deemah H. Tashman, Soumaya Cherkaoui	Trustworthy AI-Driven Dynamic Hybrid RIS: Joint Optimization and Reward Poisoning-Resilient Control in Cognitive MISO Networks	2026	Early Access	Reconfigurable intelligent surfaces Reliability Optimization	Cognitive radio networks (CRNs) are a key mechanism for alleviating spectrum scarcity by enabling secondary users (SUs) to opportunistically access licensed frequency bands without harmful interference to primary users (PUs). To address unreliable direct SU links and energy constraints common in next-generation wireless networks, this work introduces an adaptive, energy-aware hybrid reconfigurable intelligent surface (RIS) for underlay multiple-input single-output (MISO) CRNs. Distinct from prior approaches relying on static RIS architectures, our proposed RIS dynamically alternates between passive and active operation modes in real time according to harvested energy availability. We also model our scenario under practical hardware impairments and cascaded fading channels. We formulate and solve a joint transmit beamforming and RIS phase optimization problem via the soft actor-critic (SAC) deep reinforcement learning (DRL) method, leveraging its robustness in continuous and highly dynamic environments. Notably, we conduct the first systematic study of reward poisoning attacks on DRL agents in RIS-enhanced CRNs, and propose a lightweight, real-time defense based on reward clipping and statistical anomaly filtering. Numerical results demonstrate that the SAC-based approach consistently outperforms established DRL base-lines, and that the dynamic hybrid RIS strikes a superior trade-off between throughput and energy consumption compared to fully passive and fully active alternatives. We further show the effectiveness of our defense in maintaining SU performance even under adversarial conditions. Our results advance the practical and secure deployment of RIS-assisted CRNs, and highlight crucial design insights for energy-constrained wireless systems.	10.1109/TNSM.2026.3660728
Aleksandar Avdalović, Joseph Khoury, Ibrahem AlJabea, Mustafa Hajij, Elias Bou-Harb	Beyond Flat Models: Preserving ICS Topology using Combinatorial Complexes for Hierarchy-Aware Intrusion Detection	2026	Early Access	Modeling Cells (biology) Integrated circuits	Industrial Control Systems (ICS) follow the Purdue model hierarchy, organizing sensors, actuators, Programmable Logic Controllers (PLCs), and supervisory control into distinct hierarchical layers. Despite this inherent structure, existing data-driven intrusion detectors still fail to model this intricate property. To address this gap, we introduce a Hierarchy-Aware Intrusion Detector, HAID, a topology-preserving framework that models an ICS as a Combinatorial Complex (CC), operating solely on process telemetry and trained on benign traces to detect falsified sensor measurements and unauthorized actuator commands. The complex captures devices, functional interactions, and PLC control zones within an ICS. CCs (unlike flat models) encode telemetry across hierarchical levels for training a Topological Neural Network (TNN). As such, HAID enables the embedding of the ICS Purdue model into the learning process, yielding anomaly scores and localization at each of the aforementioned levels. Evaluated on two water testbeds, SWaT and WADI, and on a simulated chemical process, TEP, HAID achieves competitive detection performance (82.7% F1 on SWaT and 50.1% on WADI, averaged over five runs), while providing explicit hierarchical localization unavailable in prior work. Through ablations, we confirm the added-value of hierarchical modeling, as removing the PLC-zone representation reduces F1 by 5.8 percentage points on SWaT and causes detection on WADI to collapse. The largest losses are on attacks that manipulate several devices at once, which device- and interactionlevel signals usually miss.	10.1109/TNSM.2026.3746967
Yule Han, Kezhi Wang, Yizhe Zhao, Kun Yang	Smart-TCP: An Agentic AI-based Autonomous and Adaptive TCP Protocol	2026	Early Access	Modeling Protocols TCP	The Transmission Control Protocol (TCP) relies on deterministic state machines and sequence-number arithmetic to ensure reliable communication. However, hard-coded protocol logic is difficult to adapt to increasingly complex and dynamic network conditions. This paper proposes Smart-TCP, an agentic AI-based transport protocol framework that organizes TCP control logic as a fast/slow model-assisted decision process. Specifically, a feature-aware classifier routes routine segments to a lightweight Small Language Model (SLM) fast path and anomalous or boundary cases to a Large Language Model (LLM) slow path, while an Arithmetic Logic Unit (ALU) handles deterministic computation of sequence and acknowledgment numbers. A state module maintains connection state, packet history, and key control variables, enabling protocol decisions over the full session context. A scoped Protocol Guard checks candidate decisions against RFC-derived rules before execution. We evaluate Smart-TCP through path classification, atomic packet processing, slow-path anomaly response, and end-to-end session interaction. Experimental results show that Smart-TCP achieves 99.14% action accuracy on 700 slow-path anomaly and fallback samples, and a 100% full-lifecycle success rate over 300 ideal fast-path end-to-end session windows. The ordinary 32-KiB transfer tests achieve a 96.5% completion rate at 20% segment loss. These results suggest that separating model-based protocol decisions from deterministic arithmetic improves the reliability of TCP control for network service operation.	10.1109/TNSM.2026.3747846
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	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
Miguel Familiar-Cabero	Zero-Parameter 3GPP Protocol Prior for Infrastructure-to-CNF Event Correlation in AI-Native Mobile Core Networks	2026	Early Access	Modeling Protocols 3GPP	A zero-parameter protocol prior derived from a 3GPP Release 19 knowledge graph constrains cross-attention between infrastructure and core-network event streams to the 50 admissible (event-type, procedure-step) pairs, removing 98.6% of candidates while preserving over 99% of causal links. Evaluated on 360 scenarios spanning six 3GPP procedure families (12 scenario types) and 631,748 protocol-admissible pairs from one commercial 5G Core deployment, the method achieves leave-one-type-out cross-validated macro AUC-PR 0.823 across heldout fault types and in-sample AUC-PR skill 0.972 (0.973±0.002 over 5 seeds). On the protocol-admissible candidate set (P), the protocol-constrained neural correlation model (PCNC) outperforms an unconstrained Transformer (skill 0.901), GNN-GAT (0.935), and two gated tabular baselines (logistic regression, 0.355; gradient-boosted trees, 0.737); a no-context ablation decomposes the gap: learned sequence embeddings account for 93% of the improvement over gradient-boosted trees (Δ = +0.218), while attention-routed context adds 7% (+0.017). Controlled ablations attribute the dominant discrimination to the protocol gate (Δ Skill = +0.077, p = 0.011 cluster bootstrap) and timer bounds (+0.084), not the encoder backbone (\|Δ\| ≤ 0.006, not significant). Leave-one-EI-type-out transfer confirms that the structural prior generalizes: 12 of 14 infrastructure event types score adequately without pattern-specific supervision (mean skill 0.708). Cross-operator robustness is validated by timer perturbation (skill ≥ 0.945 at 0.5–1.0× normative values) and single-procedure KG dropout (skill ≥ 0.968). Fail-closed behavior is confirmed empirically: novel infrastructure event types absent from the KG receive exactly zero correlation score.	10.1109/TNSM.2026.3749666
Ramazan Enisoglu, Burak Alp Inan, Muhammed Ali Erbir	MSCANet: A Multi-Scale Channel Attention Network for Encrypted Traffic Classification	2026	Early Access	Modeling Transformers Convolutional neural networks	Encrypted communication has rendered traditional traffic classification ineffective, creating challenges for QoS provisioning, network management, and security monitoring. Existing deep learning approaches exploit either spatial or sequence-level features in isolation, without jointly modeling multi-scale byte patterns and inter-packet ordering. We propose MSCANet, a hybrid architecture integrating a multi-scale 1D convolutional module, a sequentially applied Efficient Channel Attention (ECA) module, and a Transformer encoder, combining the two feature streams through a learnable weighted concatenation. With approximately 1.18M parameters, MSCANet attains 95.14% mean accuracy on the 20-class USTC-TFC2016 dataset and 92.61–98.01% across three 10-class sub-tasks over five random seeds, at a median per-sample inference latency of 1.29 ms on an NVIDIA RTX 5090 Laptop GPU. Against five baselines re-trained under an identical pipeline and schedule, spanning classical machine learning, CNN-only, attention-BiLSTM, and hybrid CNN–Transformer families, MSCANet is the most accurate on all four sub-tasks, improving over the strongest classical baseline by 6.16 pp and over the closest neural baseline by 2.38 pp while using 17.5% of its parameters and running 3.1× faster. These gains are demonstrated against the five re-trained baselines. Pretrained Transformers such as ET-BERT and YaTC report higher accuracy in their original publications, under their own preprocessing and with a pretraining stage that MSCANet does not use; the advantages we expect relative to them (no pretraining stage and, against ET-BERT, a model two orders of magnitude smaller) were not measured under our protocol, and we position MSCANet within the train-from-scratch regime.	10.1109/TNSM.2026.3748386