Providing large-scale video surveillance in remote areas is challenging due to limited terrestrial communication infrastructure and the high data rates required by raw video transmission. Low Earth Orbit (LEO) satellites offer a viable connectivity solution, but their uplink capacity remains constrained, particularly for low-power devices. Semantic communication addresses this limitation by transmitting task-relevant information rather than raw video streams. In this work, we address the semantic information transmission problem in camera-to-satellite uplink communication. The task consists of selecting semantic representations from all cameras to maximize the total transmitted semantic information under strict resource constraints. We formulate the problem as an NP-hard constrained optimization task and address it using a low-complexity and scalable graph node classification approach that recasts representation selection. We propose a lightweight graph neural network that exploits the camera–satellite topology to predict the optimal semantic representation for each camera by learning context-aware node embeddings through message passing. A search-based post-processing algorithm is introduced to enforce the feasibility of the predicted solutions within transmission time requirements. Our extensive evaluations show that the proposed approach outperforms learning-based and search-based baselines in terms of accuracy, optimality, and constraint satisfaction, while maintaining low inference latency. The proposed system scales to large camera networks and enables semantic information transmission with minimal information loss and an average gap of 0.67% compared to the solver-based reference solution.
With continuous advancements in robotics and networking, the adoption of smart industrial machines for automation has grown significantly. Leveraging the Robot Operating System version 2 (ROS-2) and its topic-based data distribution mechanism, these machines are typically controlled by software deployed in the cloud. This control software is structured as a chain of ROS-2 nodes, with each node responsible for managing a specific function or component of the industrial machine. By applying the concept of Service Function Chaining (SFC), these ROS-2 node chains can be deployed in a distributed manner across cloud infrastructure. An SDN controller is employed to steer ROS-2 traffic efficiently along the sequence of nodes. As the number of industrial machines increases in large-scale factories, the number of corresponding ROS-2 node chains also grows. This trend underscores the need to reuse common ROS-2 nodes across multiple chains to optimize the utilization of cloud resources. However, the default topic-based communication model of ROS-2 is not inherently compatible with such reuse. Sharing common ROS-2 nodes across different chains becomes challenging, thereby hindering efficient resource management in cloud environments. To address this limitation, this paper proposes a Data Distribution Routing (DDR) framework powered by a Spatial Attention Convolutional Neural Network (SA-CNN). The DDR is capable of classifying and steering ROS-2 traffic originating from shared nodes to their appropriate service chains. Furthermore, it supports the NetConf protocol to communicate with the Software-Defined Networking (SDN) controller. Our evaluation includes both neural network performance and system-level metrics. The SA-CNN outperforms Vanilla CNN and CNN+LSTM in classification accuracy and efficiency. Additionally, the DDR shows favorable resource usage, latency, and throughput under real-world workloads.
The rapid growth of massive machine-type communications (mMTC), combined with advances in edge intelligence, is paving the way for low-latency, low-overhead connectivity. However, the sporadic nature of device activity in mMTC scenarios calls for efficient methods to determine which devices are active at any given time. This motivates collaborative learning within a cell-free massive multiple-input multiple-output (CF-mMIMO) architecture, where the wide geographical distribution of access points (APs) and their joint coordination make distributed learning efficient and secure. Consequently, federated learning (FL) emerges as a promising solution. Indeed, FL enables participants to train a shared model without exchanging raw local data, thereby enhancing data privacy at the AP side and lowering the fronthaul load while leveraging heterogeneous, location-dependent data. The present study proposes a novel FL framework where the CF-mMIMO participants are APs. Due to differences in device behavior, mobility patterns, and environmental factors across the network, the data collected at each AP is often non-independent and non-identically distributed (non-IID). This heterogeneity slows down the convergence of standard FL training and increase variability among client updates, particularly under heterogeneous radio feature distributions. To address this, we propose a client selection strategy that prioritizes APs based on their average received signal power. Our approach shows competitive performance compared to baseline methods, while also addressing the scalability and privacy requirements of mMTC systems. Furthermore, our study analyzes the fairness achieved by APs across devices and presents a representative, percentage-scale analysis of power-consumption gains relative to detection performance when some APs are dropped (i.e., taken out of service), examining two AP-dropping strategies. These results bring valuable insights and set guidelines towards the implementation of FL-based activity detection in CF-mMIMO networks.
Efficient and fair coexistence in unlicensed spectrum is critical for next-generation wireless networks such as 5G New Radio in Unlicensed Spectrum (NR-U) and Wi-Fi, which contend for shared resources under diverse Quality of Service (QoS) requirements. Existing reinforcement learning (RL) approaches either rely on multi-objective formulations with heuristic weight tuning or primal-dual constrained methods that struggle with stability in dynamic environments. To address these limitations, we propose QaSAL-CPM, a novel QoS-aware State-Augmented Learning framework for Coexistence Parameter Management (CPM). Building on the recently introduced concept of state augmentation, QaSAL-CPM embeds dual variables directly into the agent’s observation space, enabling real-time responsiveness to constraint violations without retraining or complex penalty tuning. This design achieves strict QoS guarantees for high-priority traffic while maintaining fairness across traffic classes, even under heavy contention. Unlike prior methods, QaSAL-CPM separates offline learning from lightweight online execution, making it practical for real-world deployments such as Ultra-Reliable Low-Latency Communications (URLLC) and the Internet of Things (IoT). Through extensive simulations of 5G NR-U/Wi-Fi coexistence scenarios, we show that QaSAL-CPM enforces 95th-percentile delay compliance and improves policy robustness. These results demonstrate that state-augmented constrained RL offers a scalable and adaptive solution for real-time coexistence optimization in dynamic wireless networks.
This paper develops a utility-privacy optimized federated learning (FL) framework that aggregates local updates transmitted over fading channels using only channel state information at the receiver (CSIR). We first establish a rigorous convergence analysis of distributed stochastic gradient descent (SGD) under arbitrary aggregation weights and channel fading, deriving an upper bound that explicitly captures the combined effects of channel quality and SGD noise. Using this bound, we formulate and solve an optimization problem that yields closed-form or efficiently computable aggregation weights. The resulting method provably achieves the standard O(1/t) convergence rate for strongly convex objectives, even under arbitrarily varying fading channels. We further characterize the differential privacy (DP) induced by channel noise and develop a principled mechanism that enables explicit utility-privacy tradeoffs through weight selection. Extensive experiments on MNIST, CIFAR-10, Shakespeare, and F-EMNIST demonstrate that the proposed fading-aware aggregation eliminates the training failures suffered by FedAvg under deep fading and approaches the performance of ideal error-free communication. The results provide a unified convergence-communication-privacy framework for robust and privacy-aware FL over wireless fading channels.
FL intrusion detection systems (IDSs) have emerged as a defense-in-depth layer for securing internet of things (IoT) networks by training detection models and classifying IoT traffic locally on distributed devices. However, developing effective federated learning (FL)-IoT IDSs remains significantly challenging due to key IoT characteristics, including their distributed nature, resource constraints, and data heterogeneity. Thus, improving detection performance and accelerating model convergence in FL-IoT IDSs while accounting for these IoT characteristics is essential. Moreover, existing FL-IoT IDSs lack mechanisms for preparing reliable pseudo-labeled IoT traffic for subsequent training rounds. To this end, we propose a multi-temporal device clustering FL-IoT IDS framework that partitions federated learning detection rounds (FLDRs) in time and clusters devices within each partition to enhance detection effectiveness without compromising IoT resources. We also establish the FLDR as an organizational framework to structure the periodic intrusion-detection processes in FL-IoT IDSs. The framework employs a post-processing (PP) module that applies confidence thresholding and outlier removal in unsupervised and hybrid supervised-unsupervised settings, enabling on-device filtering of pseudo-labeled traffic and reuse of retained pseudo-labeled traffic in the subsequent round. It also incorporates a data-driven clustering technique that anticipates non-independent and identically distributed (IID) behavior in upcoming FLDRs, thereby mitigating the impact of non-IID intrusion distributions. For evaluation, we implement dual-, triple-, and quadruple-temporal variants of the framework and assess different machine learning components. The PP module is benchmarked against supervised and no-PP baselines, while the data-driven clustering technique is compared with random and similarity-based clustering approaches. The data-driven quadruple-temporal clustering technique is also benchmarked against two existing FL-IoT IDS algorithms. We further assess resource consumption in terms of computational usage and communication overhead. Experiments on balanced and imbalanced intrusion scenarios derived from the Edge-IIoTset and MQTT-IoT-IDS datasets demonstrate that the proposed framework achieves superior detection accuracy, faster convergence, and fewer false positives and false negatives, thereby reducing unnecessary alarms and lowering the risk of undetected intrusions. The PP module proves effective for on-device pseudo-label filtering, with hybrid supervised-unsupervised methods achieving performance close to that of the supervised baseline, while the data-driven clustering technique consistently yields balanced and robust performance. Moreover, the proposed data-driven quadruple-temporal variant outperforms two existing FL-IoT IDS algorithms on imbalanced datasets. Importantly, the overall resource utilization remains comparable to that of the traditional FL-IoT IDS approach.
As wireless communication technology advances toward sixth-generation (6G) systems, intelligent networks are evolving into user-centric, decentralized, and highly heterogeneous environments. Such systems’ multi-source heterogeneous time series data exhibit unprecedented scale, complexity, and strong spatiotemporal coupling. Challenges commonly encountered in traditional domains—such as complex temporal dependencies, high dimensionality, and scarcity of labeled anomalies—become more pronounced in 6G scenarios and are further compounded by intricate spatial topologies and dynamic inter-entity relationships. To address these challenges, this paper proposes a Knowledge Graph-enhanced Spatiotemporal Dual-path Anomaly Detection framework (KG-STDAD). The framework integrates community-level semantic information from a Knowledge Graph (KG) with a Transformer architecture with distinct temporal and spatial processing paths enhanced by variational autoencoder (VAE)-based reconstruction capabilities, enabling joint modeling of spatiotemporal dependencies. Specifically, a Transformer-VAE module is designed to capture dynamic temporal patterns. At the same time, community ID embedding, community depth masking, and edge encoding in the attention are introduced to enhance spatial modeling guided by structural knowledge from communication KGs. Extensive experiments on three datasets demonstrate the superior performance of KG-STDAD. Ablation studies further confirm the key contributions of the dual-path architecture and knowledge graph-based encoding. This work provides an accurate and explainable anomaly detection solution for user-centric 6G networks and new insights into the fusion of knowledge-driven and data-driven methods.
Orchestrating services across heterogeneous 6G edge-cloud infrastructures requires autonomous coordination systems managing distributed computational resources while satisfying Quality-of-Service (QoS) requirements. Recent advances in Large Language Models (LLMs) enable development of autonomous agents capable of complex reasoning and decision-making for such orchestration tasks. However, applying generic agentic AI frameworks from the machine learning literature to orchestration domains introduces reliability limitations, as trial-and-error decision patterns are unsuitable for environments where errors disrupt services. This work presents AgentEdge, a novel distributed intelligence framework that implements specialized autonomous agents in four orchestration roles: intent processing, infrastructure monitoring, strategic planning, and action execution. AgentEdge introduces the PARES (Perceive, Act, Reason, Evaluate, Sustain) framework establishing minimum capabilities required for autonomous agent qualification. Central to AgentEdge is the ActSimCrit (Action-Simulation-Critic) planning methodology, which validates orchestration plans through digital twin simulation before execution, eliminating direct infrastructure experimentation risks. Agents coordinate multi-step operations and adapt strategies based on constraint feedback. Structured outputs constrain agent decision spaces to feasible orchestration actions while preserving optimization flexibility. Experimental evaluation in six orchestration scenarios validates AgentEdge through comparison with baseline agentic frameworks and ablation studies. AgentEdge achieves 2.76 & times; higher success rate compared to generic agentic frameworks (ReAct, LATS) and 10 & times; reduction in API call variability. The core ActSimCrit digital twin component alone contributes 1.47 & times; success improvement when compared to direct planning without simulation. AgentEdge achieves significant power savings across infrastructure scales from 8 to 35 nodes.
The emergence of different traffic types with diverse requirements, such as eXtended Reality (XR), in 5G-Advanced and upcoming 6G networks, requires more intelligent duplexing strategies. Sub-band full duplex (SBFD) enables simultaneous uplink (UL) and downlink (DL) transmissions over an unpaired spectrum, thereby providing more frequent UL opportunities. While SBFD has strong potential to improve UL performance without risking DL performance, static radio frame configurations are inherently suboptimal under dynamic traffic conditions. In such cases, dynamic radio frame configurations are required to adapt to varying traffic demands and service requirements and fully realize the benefits of SBFD. This paper proposes an ML-enabled dynamic radio frame selection framework based on a Soft Actor-Critic (SAC) reinforcement learning agent with a two-step bootstrap mechanism. A bootstrapped general model is trained across varied traffic loads and packet sizes to ensure scalability and robustness. System-level simulations in dense urban deployments demonstrate that the bootstrap generalized model achieves good performance, while maintaining adaptability to unseen traffic conditions. Compared with conventional static time-division duplexing (TDD) baseline with “DDDSU”, the proposed scheme increases UL capacity by up to 268% for large-packet file transfer protocol model 3 (FTP3) traffic and 195% for small packets. For XR services, the ML-based solution extends UL capacity by 25%, confirming its effectiveness in supporting latency-sensitive traffic scenarios. The simulation results further indicate that SBFD can deliver a high satisfaction ratio for XR users by ensuring sufficient throughput for cell-edge users.
Internet of Things (IoT) systems often host services with diverse and sometimes conflicting Quality of Service (QoS) requirements, presenting a significant challenge in delivering tailored QoS. This challenge becomes even more complex with the introduction of autonomy, which aims to minimize or eliminate human intervention while still maintaining QoS. Network Slicing has emerged as a promising solution by enabling the creation, management, and termination of logical networks known as network slices, each designed to meet the specific QoS needs of individual IoT services. The central question is: how can we autonomously create, manage, and terminate the maximum number of slices in dynamic environments? Deep Reinforcement Learning (DRL) agents have partially addressed this challenge by autonomously and adaptively placing slices using reward-based feedback mechanisms. However, for a DRL agent to be truly effective in Network Slicing, it must meet four key criteria: fast convergence, high slice placement performance, policy stability and strong scalability to maintain the first three qualities regardless of the underlying physical infrastructure’s size. To the best of our knowledge, no existing solution satisfies all four criteria. In this work, we present DAAFA, a DRL agent specifically designed to meet this challenge in Core Network slicing. DAAFA achieves high scalability by decoupling convergence speed from infrastructure size and ensures rapid, stable convergence through advanced feature extraction techniques. To validate its effectiveness, we developed two versions of DAAFA based on Deep Q-Networks (DQN) and Double Deep Q-Networks (DDQN), and compared them against ten state-of-the-art alternatives, each tuned for optimal performance. Simulation results across three infrastructures, scaling up to 360 nodes, confirm that DAAFA successfully meets all four essential criteria.
Non-geostationary satellite orbit (NGSO) systems are expected to play a central role in enabling direct satellite-to-vehicle (DS2V) connectivity. However, dense spectrum reuse and the coexistence of multiple NGSO constellations introduce severe co-frequency interference (CFI), which is further exacerbated by fast satellite motion and short channel coherence times. In this paper, we propose a federated learning (FL)–enabled collaborative receive beamforming framework for DS2Vs (FedSatBF). In FedSatBF, a serving satellite aggregates locally trained beamforming models from a subset of vehicular user terminals (VUTs) via federated averaging and broadcasts a global model that can be directly applied by both participating and non-participating VUTs under similar CFI conditions. An attention-based beamforming architecture is adopted as the base model for training local models at each VUT, with estimation-free sample covariance matrices as input, thereby eliminating the need for explicit estimation of the interference channel or the direction of arrival. The local model is trained via a reward-based learning objective in which the loss is the negative measured post-beamforming signal-to-interference-plus-noise ratio (SINR), so that no ground-truth weight, no synthetic geometry-derived label, and no interferer’ channel state information (CSI) is provided or needed during training. The proposed framework is explicitly designed to operate within the limited channel coherence time imposed by NGSO mobility. We investigate the impact of key FL hyperparameters, including per-coherent block communication rounds, per-round training data, per-VUT local epochs, VUT participation, and geographical VUT clustering, on convergence speed, interference nulling capability, and computational efficiency. Simulation results demonstrate that FedSatBF achieves near-optimal signal-to-interference-plus-noise ratio (SINR) performance, within 0.1 [dB] of an optimal zero-forcing beamformer (ZFBF) used here as a perfect-CSI oracle upper bound, and significantly outperforms conventional maximum-ratio combining (MRC), all while using only a moderate FL hyperparameter setting. Moreover, the global model exhibits strong generalization to non-participating VUTs and remains effective under reduced settings, highlighting the practicality of FL-based collaborative beamforming for the next generation of DS2V systems.
User selection is an important component of the Radio Resource Management of fully digital Multi-User Multiple-Input Multiple-Output (MU-MIMO) systems and its efficiency is critical to the performance of these systems. In this paper, we propose a novel framework called, Dynamic Probabilistic User Selection (DPUS), that formulates user selection as a probabilistic inference task using Markov Random Fields and Loopy Belief Propagation. Unlike existing methods, DPUS operates in an online unsupervised manner, eliminates the need for labeled data, and integrates fairness constraints with channel-awareness. Importantly, DPUS can apply to both the downlink and the uplink. Extensive simulations conducted in a large range of deployment scenarios show that DPUS consistently outperforms state-of-the-art heuristics in terms of fairness, scalability, and runtime. DPUS provides a robust solution for the efficient operation of fully digital MU-MIMO systems, paving the way for next-generation wireless networks.
Neural decoders have recently advanced error correction performance, with transformer-based architectures in the error correction code transformer (ECCT) framework achieving state-of-the-art results among neural decoders. However, the multiple-masks ECCT suffers from suboptimal mask selection, which impacts decoding performance. In this work, we propose spectral-guided ECCT, leveraging the Laplacian matrix and its eigenvectors from the variable-variable graph to identify weakly connected node clusters for optimized mask generation. By embedding spectral properties into the decoding process, our method constructs more effective masks that improve information flow across variable and check nodes. Experimental results demonstrate consistent gains over existing ECCT-based models, achieving superior decoding accuracy and reliability while significantly improving training efficiency. This work highlights the potential of spectral graph theory to address structural challenges in neural decoders and provides a principled direction for future error correction design.
Wireless communication technology is a vital enabler for 5G vehicle-to-everything (5G-V2X) communications, which are vulnerable to cyberattacks such as distributed denial of service (DDoS). DDoS detection requires sophisticated artificial intelligence (AI)-based approaches, whose main challenge is data privacy. Secured peer-to-peer (P2P) federated learning (FL) techniques could overcome this challenge by training detection models locally while sharing model parameters for secure aggregation. However, such a strategy scales poorly and incurs a high communication cost as the number of FL clients grows. In such a context, we propose SAFE-ADVENT, a secure P2P FL strategy that reduces the communication cost, incorporating transfer learning and client selection within the P2P FL system to detect DDoS attacks. By using data collected from our 5G-V2X testbed, under offline scenarios, SAFE-ADVENT converged faster than deep learning, centralized FL and SAC-based P2P FL benchmarks. Through live traffic testing and under adversarial attacks, SAFE-ADVENT outperformed all benchmarks in most scenarios. Finally, SAFE-ADVENT proved to be scalable by preserving high performance when increasing the number of FL agents.
Deep learning-based wideband spectrum sensing models have achieved impressive performance in subband occupancy detection, but their susceptibility to adversarial attacks raises serious concerns. In this paper, we construct a targeted adversarial attack framework specifically designed for multi-label wideband spectrum sensing tasks. We further propose two optimization-based attack methods. MLS CW formulates perturbation generation as a constrained optimization problem and solves it via gradient descent. MLS DF is a greedy iterative method that approximates the solution to an underdetermined linear system through linearization. Both attacks are developed under strict power constraints and are capable of generating subtle yet highly effective perturbations. These attacks further incorporate a threshold-aware loss and selective subband targeting to enhance precision and stealthiness. Extensive experiments across various SIR conditions demonstrate that the proposed attacks significantly degrade the model’s detection performance in both untargeted and targeted scenarios, while maintaining high stealthiness. These results expose the severe vulnerability of DL-based spectrum sensing models and highlight the need for further research into their robustness under adversarial settings.
This paper investigates a dual-function radar-communication (DFRC) system, which senses a malicious target and communicates with legitimate users simultaneously. Being aware of the possibility that the target can overhear confidential messages inherently intended for the users, the DFRC beampattern design incorporates geofencing considerations, particularly in the presence of target mobility. To this end, we design adaptive beampatterns based on Q-learning, which dynamically adjusts the beampattern to accommodate target mobility, communication, and geofencing constraints. However, under such constraints, the transmit beampattern may become suboptimal for sensing functionality due to angular mismatches between the target direction and the feasible beampattern direction. To overcome this issue, we employ Gaussian Process Regression (GPR) to infer the target’s angle of arrival, even when the transmit beam is misaligned with the target (e.g., when the target enters geofenced regions). The combination of Q-learning and GPR provides a flexible machine learning-based framework for geofencing-aware DFRC systems in dynamic and constrained scenarios.
Traffic routing is crucial for the efficient operation of the Internet. As network traffic grows, adaptive and intelligent routing algorithms are needed to meet various Quality of Service (QoS) requirements. Reinforcement Learning (RL)-based routing approaches have demonstrated superior performance over traditional methods. In this study, we propose RL-based Segment Routing (SR), called RLSR, a QoS-aware routing algorithm for Software-Defined Networking (SDN). Our approach ensures loop-free path exploration and leverages learned network QoS states to accelerate convergence when optimizing paths for multiple traffic demands. By integrating SR, RLSR enables flow-based source packet routing while reducing communication overhead between the SDN controller and network switches. Experimental evaluations demonstrate that RLSR improves load balancing by reducing the maximum link utilization from 71.67% to 50% compared with a non-RL baseline. In addition, the proposed knowledge-reuse mechanism accelerates learning, reducing the total number of training episodes by up to 27% compared with a baseline RL configuration without knowledge reuse. Further analysis shows that a higher discount factor ( gamma=0.9) ) leads to the fastest convergence among the tested settings. These findings highlight the effectiveness of RLSR in optimizing traffic engineering and network performance in dynamic environments.
This paper introduces a proactive Unmanned Aerial Vehicle (UAV) mobility management xApp for Open Radio Access Network (O-RAN) Near Real-Time Radio Intelligent Controller (Near-RT RIC) environments, employing Double Deep Q-Network (DDQN) reinforcement learning (RL) enhanced with transfer learning to optimise handover decisions for UAVs operating along predetermined flight trajectories. Unlike reactive approaches that respond to signal degradation, the proposed framework anticipates network conditions and minimises both outage probability and handover frequency through predictive optimisation. The system leverages centralised weight averaging to consolidate knowledge from multiple flight scenarios into a global model capable of generalising to previously unseen operational environments without extensive retraining. A comprehensive evaluation demonstrates that the proposed framework achieves a favourable trade-off between handover frequency and connectivity reliability, reducing handover events by up to 54.6% compared to greedy approaches while maintaining outage probability at practically negligible levels. The results validate the effectiveness of intelligent learning-based approaches for UAV mobility management in next-generation O-RAN architectures, thereby contributing to seamless integration of aerial user equipment into cellular networks.
Digital twins of networked infrastructures, known as Virtual Infrastructure Twins (VITs), are increasingly used for software development, pre-deployment testing, and design space exploration. While VITs avoid the costs and potential disruptions associated with experiments on operational networks, their throughput measurements are typically not sufficiently accurate for performance profiling of wide-area networks that they emulate. Here, machine learning (ML) methods are developed to transform these inaccurate VIT network throughput measurements to closely match in peak and overall profile of those from a physical testbed or production network. First, a micro kernel network reflecting a physical network is utilized to collect one-time measurements on a host to support this ML transformation. Then, a generic multi-modal ML method is developed to learn a map that transforms measurements from subsequent VITs on the same host to match past, current and follow-on testbed and cloud networks. ML generalization equations are derived to establish its correctness and probabilistically guarantee its generalization accuracy. Experimental results are presented for a variety of VIT hosts with target testbed and cloud networks; they include a case study of a four-site science ecosystem wherein inaccurate convex VIT measurement profiles are transformed into accurate concave profiles of target networks.
