Open Radio Access Network (O-RAN) disaggregation introduces architectural flexibility but significantly expands the cellular attack surface. Specifically, Radio Resource Control (RRC) signaling storms exploit these functional splits, severely amplifying resource exhaustion across distributed RAN components. State-of-the-art defenses typically deploy security rApps or xApps at the core or upper network layers; however, these approaches are inherently reactive, detecting anomalies only after radio resources and control-plane entities are already saturated. By relying on aggregated counters and volumetric thresholds, they overlook the immediate physical contention and disruption occurring at the radio edge. To address these limitations, we propose RadioShield, a novel, multi-tier defense framework that shifts the detection scope from upper-layer monitoring to the PHY/MAC layer at the Open Distributed Unit (O-DU). By leveraging radio telemetry, RadioShield enables early detection through a distributed application (dApp), in close proximity to malicious User Equipment (UEs), allowing the network to identify and halt signaling storm traffic microseconds after it originates. To validate this framework, we utilized a physical O-RAN testbed with realistic UE reboot and handover storms, demonstrating over 99% accuracy in both binary and multi-class detection. Computational profiling confirms microsecond-scale inference and reduced resource overhead, proving that dApp-based early detection outperforms traditional, upper-layer network monitoring in both computational efficiency and threat mitigation speed.

Explainable AI (XAI) has gained prominence in Intrusion Detection Systems (IDS) to address the opacity of black-box models. However, existing efforts have largely confined XAI to post-hoc interpretability, with limited exploration of its potential for driving data refinement and model optimisation. To the best of our knowledge, SHAPRefine is the first model-agnostic feature selection framework that systematically integrates class-wise SHAP value aggregation with automated refinement to produce lightweight and high-performing IDS models. Unlike existing SHAP-based methods that repeatedly recompute Shapley values for each candidate subset—often requiring excessive retraining and parameter tuning—SHAPRefine introduces a performance-driven iterative selection strategy that minimizes redundant computations while explicitly accounting for minority-class contributions. Comprehensive evaluations across diverse IDS datasets and classifiers demonstrate that SHAPRefine consistently reduces feature dimensionality by up to 90%, cuts training time and energy consumption, and improves detection accuracy across all attack types, with F1-score gains of up to 95% for minority classes. These results establish SHAPRefine as a robust and resource-efficient solution for accurate IDS deployment in Beyond 5G (B5G) and 6G networks.

With the rise of next-generation wireless systems, Deep Learning (DL)-based security solutions are gaining widespread adoption. One key area DL impacts is beamforming prediction in mmWave communication. This study investigates how vulnerable AI-based models are to adversarial attacks, which can disrupt network performance in complex distributed MIMO (D-MIMO) environments. Since AI plays a crucial role in selecting beams efficiently, such vulnerabilities pose a serious threat to the reliability of 5G and 6G networks. Existing work fails in achieving generalizable performance on unseen attacks in low-latency requirements. To enhance resilience, we propose an innovative XAI-based framework dubbed X-ZeroSec that leverages SHAP, specifically Shapley values from cooperative game theory, to detect and mitigate adversarial attacks. Our approach introduces the concept of explanation distillation, where feature attributions from a “teacher” detector model train a generalized XAI-based “student” model, yielding improved and generalized detection rates, especially for zero-shot attacks. The framework enhances robustness against various adversarial patterns, achieving over 10-16% improved accuracy compared to state-of-the-art detectors. Comprehensive evaluations demonstrate that integrating XAI into AI-driven beamforming fortifies model defenses, preserving mmWave system performance while meeting evolving security requirements. X-ZeroSec framework supports diverse DL-based applications and it is implemented on the UCD NETSLAB 5G O-RAN testbed, demonstrating promising end-to-end latency of less than 4 milliseconds.

As Open Radio Access Networks (O-RAN) accelerate 5G deployments through intelligent, software-defined architectures, they concurrently expand the cellular attack surface. Deploying and validating Intrusion Detection Systems (IDS) against attacks such as Radio Resource Control (RRC) signaling storms and volumetric network floods requires scalable, realistic testbeds. In this demonstration, we present a software-simulated O-RAN environment capable of multi-UE connectivity and executing coordinated multi-layer attacks. By leveraging srsRAN, Open5GS, FlexRIC, and a custom GNU Radio ZeroMQ multiplexer, we demonstrate a dual-layer defense framework featuring a real-time Monitoring xApp that bridges physical telemetry and network flows to AI-driven IDS xApps, enabling the rapid detection of malicious activities within the Near-Real-Time RAN Intelligent Controller (Near-RT RIC).

Deploying Large Language Models (LLMs) in edge computing environments is frequently hindered by stringent memory constraints, high computational demands, and strict privacy regulations. While distributed paradigms like Split-Federated learning mitigate privacy risks by computing partial forward passes locally, executing transformer layers on resource-constrained devices remains a computational bottleneck. This paper presents a functional demonstration of TITAN, a novel hardware-software co-design framework that integrates Field-Programmable Gate Array (FPGA) acceleration into the Distributed LLM architecture. By leveraging a hybrid CPU-FPGA execution model for parameter-efficient fine-tuning, TITAN provides a scalable, energy-efficient solution for privacy-preserving LLM workloads at the network edge.

The Sixth-Generation (6G) are already in the horizon, owing to advents of communication technologies towards enabling intelligent applications and services. Federated Learning (FL) is a distributed Artificial Intelligence (AI) technology that underpins 6G communication technologies and applications. Interestingly, FL is also a promising contender to enhance 6G security. This paper presents a comprehensive and up-to-date review of FL-enabled 6G security. The paper explores security threats in FL for 6G, threats in FL for 6G, and threats shared across FL and 6G. Subsequently, how FL can be utilized to strengthen 6G security in the Radio Access Network (RAN), Open RAN (O-RAN), network edge, and network orchestration and core is presented. In addition, FL is for 6G application and service security across various emerging applications, ranging from Connected Autonomous Vehicles (CAVs) to the envisaged metaverse applications. The paper then consolidates lessons learned, projects, and proposes future research directions to establish the role of FL in strengthening 6G security.

Harmful Algae Blooms (HABs) pose significant threats to water quality, ecosystem health, and human safety, resulting in substantial economic losses in aquaculture and tourism. Existing HAB prediction methods often rely on singlemodality data and are limited to short prediction horizons of less than one week, making them struggle to capture the complex spatio-temporal environmental drivers of bloom dynamics and to provide sufficient lead time for proactive coastal management. This paper proposes HABFusion, a multimodal fusion framework that uses a hybrid fusion strategy to integrate heterogeneous data modalities from satellite imagery, in-situ observations, and meteorological models for multi-step HAB prediction. By jointly predicting multiple future prediction horizons within a unified model, the framework improves efficiency and prediction consistency compared to training separate models for individual horizons. By learning cross-modal dependencies, HABFusion consistently outperforms image-only baselines across prediction horizons from 1 to 14 days ahead, achieving an average F1-score improvement of 8% in a case study of Coastal Florida. The code is publicly available at: https://github.com/Guojyjy/habfusion.

Dynamic Spectrum Sharing (DSS) enables flexible and efficient utilization of spectrum resources, but existing Non-Fungible Token (NFT)-based frameworks suffer from critical privacy limitations. Although public blockchains provide pseudo-anonymity by masking direct identities, the transparency of transaction histories allows adversaries to trace user activities, link addresses, and infer spectrum usage patterns, exposing spectrum owners to targeted attacks and misuse of idle bands. This weakness undermines the security, fairness, and trustworthiness of NFT-enabled DSS. To mitigate these vulnerabilities, Anon-Spect is introduced as a privacy-enhanced NFT-based DSS framework that employs ephemeral keys, Zero-Knowledge Proofs (ZKPs), Pedersen commitments, and linkable ring signatures to achieve anonymity, unlinkability, and metadata obfuscation while maintaining on-chain verifiability. An Ethereum-based implementation demonstrates that Anon-Spect achieves near-zero spectrum misuse and strong resilience against collusion attacks while preserving market efficiency. Although execution costs are higher compared to baseline frameworks, the overhead scales linearly and remains practical relative to the strengthened privacy and security guarantees.

With the advent of 5G commercialization, the need for more reliable, faster, and intelligent telecommunication systems is envisaged for the next generation beyond 5G (B5G) radio access technologies. Artificial Intelligence (AI) and Machine Learning (ML) are immensely popular in service layer applications and have been proposed as essential enablers in many aspects of 5G and beyond networks, from IoT devices and edge computing to cloud-based infrastructures. However, existing 5G ML-based security surveys tend to emphasize AI/ML model performance and accuracy more than the models’ accountability and trustworthiness. In contrast, this paper explores the potential of Explainable AI (XAI) methods, which would allow stakeholders in 5G and beyond to inspect intelligent black-box systems used to secure next-generation networks. The goal of using XAI in the security domain of 5G and beyond is to allow the decision-making processes of ML-based security systems to be transparent and comprehensible to 5G and beyond stakeholders, making the systems accountable for automated actions. In every facet of the forthcoming B5G era, including B5G technologies such as ORAN, zero-touch network management, and end-to-end slicing, this survey emphasizes the role of XAI in them that the general users would ultimately enjoy. Furthermore, we presented the lessons from recent efforts and future research directions on top of the currently conducted projects involving XAI.

Urban intersections often struggle to balance traffic efficiency with the safety of Vulnerable Road Users (VRUs) such as pedestrians. While traditional traffic light control (TLC) methods focus on optimizing vehicle flow, they often neglect pedestrian safety. Therefore, we propose SafeCrossLight, a deep reinforcement learning (DRL) - based approach that aims to address both efficiency and safety in a unified framework. By factoring pedestrians’ safety into the learning process, Safe-CrossLight enables a more responsible and adaptive decisionmaking process at intersections. Our method is evaluated through Simulation of Urban Mobility (SUMO) and compared with several state-of-the-art TLC approaches. Results show that SafeCrossLight significantly reduces unsafe pedestrian behaviors while maintaining strong efficiency in both vehicle and pedestrian flow, suggesting its high potential for real-world deployment in urban traffic systems.

As mission-critical (MC) services such as Unmanned Aerial Vehicles (UAVs) based emergency communication and Internet of Vehicles (IoVs) enabled autonomous driving emerge, the traditional communication framework can not meet the growing demands for higher reliability and lower latency and the increasing transmission loads. Semantic Communication (SemCom), an emerging communication paradigm that shifts the focus from bit-level data to its context and intended task at the receiver (i.e., semantic level), is envisioned to be a key revolution in Sixth Generation (6G) networks. However, an explicit and systematic SemCom framework specifically tailored for Vehicle-based MC (VbMC) services has yet to be proposed, primarily due to the complexity and lack of analysis on their MC characteristics. In this article, we first present the key information-critical and infrastructure-critical vehicle-based services within the SemCom framework. We then analyze the unique characteristics of MC services and the corresponding challenges they present for SemCom. Building on this, we propose a novel SemCom framework designed to address the specific needs of MC services in vehicle systems, offering potential solutions to existing challenges. Finally, we present a case study on UAV-based rapid congestion relief, utilizing eXplainable AI (XAI) to validate the effectiveness of the proposed SemCom framework.

With the expanding use of unmanned aerial vehicles (UAVs) across various fields, efficient obstacle avoidance has become increasingly crucial. This UAV obstacle avoidance can be achieved through deep reinforcement learning (DRL) algorithms deployed directly on-device (i.e., at the edge). However, practical deployment is constrained by high training time and high inference latency. In this paper, we propose methods to improve DRL-based UAV obstacle avoidance efficiency through improving both training efficiency and inference latency. To reduce inference latency, we employ input dimension reduction, streamlining the state representation to enable faster decision-making. For training time reduction, we leverage transfer learning, allowing the obstacle avoidance models to rapidly adapt to new environments without starting from scratch. To show the generalizability of our methods, we applied them to a discrete action space dueling double deep Q-network (D3QN) model and a continuous action space soft actor critic (SAC) model. Inference results are evaluated on both an NVIDIA Jetson Nano edge device and a NVIDIA Jetson Orin Nano edge device and we propose a combined method called FERO which combines state space reduction, transfer learning, and conversion to TensorRT for optimum deployment on NVIDIA Jetson devices. For our individual methods and combined method, we demonstrate reductions in training and inference times with minimal compromise in obstacle avoidance performance.

Deep Reinforcement Learning (DRL) for traffic light control adaptively adjusts signals based on real-time traffic conditions to alleviate urban congestion. However, transmitting image data from intersection cameras results in high communication overhead and latency in practical deployments. Traditional DRL methods lack interpretability when selecting some common traffic features (e.g., traffic densities, speed) as state input. Inspired by 6G semantic communication that transmits semantic information rather than raw image data, this paper proposes xTL, an eXplainable AI (XAI)-guided DRL development approach to achieve semantic state design and address communication overhead. Unlike traditional DRL that heavily depends on empirical tuning, xTL allows human experts to efficiently guide and interpret model design by using XAI-generated explanations to distill lightweight semantic traffic features from image data. Utilizing SHapley Additive exPlanations (SHAP)-generated saliency maps, we identify a new critical feature: the location of the last vehicle in the first platoon on each incoming road, which can interpret intersection traffic dynamics and effectively improve traffic. Experiments on two urban intersections in SUMO demonstrate that xTL slashes communication costs by over 90% and shortens DRL training duration by about 21%, without compromising traffic control effectiveness.

Cloud computing has revolutionized the provisioning of computing resources, offering scalable, flexible, and on-demand services to meet the diverse requirements of modern applications. At the heart of efficient cloud operations are job scheduling and resource management, which are critical for optimizing system performance and ensuring timely and cost-effective service delivery. However, the dynamic and heterogeneous nature of cloud environments presents significant challenges for these tasks, as workloads and resource availability can fluctuate unpredictably. Traditional approaches, including heuristic and meta-heuristic algorithms, often struggle to adapt to these real-time changes due to their reliance on static models or predefined rules. Deep Reinforcement Learning (DRL) has emerged as a promising solution to these challenges by enabling systems to learn and adapt policies based on continuous observations of the environment, facilitating intelligent and responsive decision-making. This survey provides a comprehensive review of DRL-based algorithms for job scheduling and resource management in cloud computing, analyzing their methodologies, performance metrics, and practical applications. We also highlight emerging trends and future research directions, offering valuable insights into leveraging DRL to advance both job scheduling and resource management in cloud computing.

Federated learning (FL), propelled by advancements in artificial intelligence and edge computing, is increasingly employed in privacy-sensitive intelligent transportation systems (ITS). However, accurately predicting the estimated time of arrival (ETA) for long vehicular trips spanning multiple regions remains challenging due to heterogeneous traffic patterns and insufficient local data. Although single-region or centralized solutions offer higher accuracy, the former struggles to address complex intra-regional dynamics, while the latter raises significant privacy concerns by aggregating large volumes of mobility data from various regional authorities by a single giant entity(e.g., Google or Alibaba). To address these challenges, we introduce a novel Multi-Region Federated Learning (MRFL) framework to collect traffic data at each region-specific base station (BS) to predict the ETA of vehicles without sharing the collected data among traffic BSs, ensuring privacy and alleviating local data scarcity. Experimental evaluations on SUMO datasets demonstrate that MRFL significantly outperforms single-region learning in prediction accuracy and convergence speed, highlighting the efficacy of MRFL in enhancing ETA prediction in diverse traffic scenarios and offering a promising avenue for future advancements in ITS.

As Federated Learning (FL) becomes a cornerstone for privacy-preserving Artificial Intelligence (AI) across mobile and edge environments, its growing energy demands pose a serious challenge to scalability, sustainability, and real-world adoption. While FL avoids centralized data sharing, its iterative training rounds, communication-heavy protocols, and added security layers significantly strain device and network resources. Recent research has introduced energy-aware techniques such as pruning methods, quantization, adaptive client selection, and lightweight secure aggregation, yet many of these solutions optimize only isolated components of the FL pipeline, often trading off model performance, fairness, or security. Despite the growing interest in energy optimization, there is currently a lack of dedicated survey work that focuses specifically on energy-efficient strategies in FL, which this paper aims to address. This paper presents a compact yet comprehensive survey of state-of-the-art energy-efficient FL strategies, systematically analyzing how they address key energy consumption hotspots in FL architecture. We map existing solutions to these components, identify their limitations, and highlight emerging directions such as role-aware energy optimization, cross-layer optimization strategies, and the paper concludes with a call for standardized benchmarking frameworks to advance FL toward a truly sustainable and deployable future.

The Internet of Things (IoT) connects numerous heterogeneous devices that collect and generate substantial sensitive data to train intelligent models. To protect the privacy of such data, federated learning (FL) provides a privacy-preserving distributed machine learning approach. However, these IoT devices are highly heterogeneous, differing in computing and communication capabilities. During the FL training, these variations may result in significant communication delays and computational inefficiencies. We propose federated fleet (FedFleet), an efficient FL framework with a novel clustering method, to accelerate model training in heterogeneous IoT device scenarios. Unlike the existing FL clustering method, which first determines the cluster header, FedFleet clusters devices into different fleets corresponding to the number of edge servers. Specifically, FedFleet divides devices into multiple fleets based on their performance profiles, including computation and communication time. Devices in the same fleet synchronously aggregate their local model parameters at the edge server and then asynchronously update the global model to the central server. Experiments with various models and datasets show that FedFleet outperforms the state-of-the-art FL algorithm, reducing convergence time by 10.2%-56.4% while achieving the same accuracy rate. It also exhibits minimal performance fluctuations under varying device heterogeneity (DH) levels, dropout rates, and network conditions. Moreover, FedFleet demonstrates better scalability for high DH environments.

In the rapidly evolving landscape of networking and security, the adoption of artificial intelligence (AI) is accelerating to meet the demands of real-time, data-driven applications. Current AI development processes predominantly prioritize model utility metrics such as accuracy, precision, and recall, often overlooking critical trustworthiness aspects like accountability, resilience to adversarial attacks, and privacy. To address this gap, we propose a novel AI/Machine Learning (ML) development process that systematically integrates trustworthiness metrics alongside traditional model utility measures. Our process emphasizes the iterative development of trustworthy AI models by balancing performance, accountability, resilience, and privacy through the incorporation of eXplainable AI (XAI) techniques. We validate the effectiveness of our methodology across four distinct networking and security use cases. In encrypted traffic classification, LightGBM emerges as the most practical model, offering a strong balance of utility, accountability, and robustness despite Neural Networks achieving the highest raw performance. For malware detection, feature reduction in the MalDoc model yields a minimal utility loss (<0.7%) while substantially enhancing resilience to evasion attacks (10–80%). In assessing privacy trade-offs in Federated Learning, we observe that although strong differential privacy significantly degrades utility (up to 70% on MNIST), it enables early-stage privacy protection without fully masking poisoned client behaviour, which remains detectable through SHAP and t-SNE-based analysis. Lastly, in a smart healthcare emergency e-call scenario, our 1D CNN model achieves not only strong predictive performance (96.21% accuracy, 91.55% precision, 93.13% recall) but also provides stable and interpretable explanations using LRP and SHAP, with LRP demonstrating higher consistency across ECG segments. Therefore, unlike prior studies that focus on isolated aspects such as accountability or resilience, our work proposes a holistic, quantifiable process that balances the trade-offs among model utility, accountability, resilience, and privacy to support the development of trustworthy AI models in communication systems.

Software-defined networking (SDN) has revolutionized modern mobile networks by enhancing flexibility and scalability, but its centralized architecture remains a prime target for Distributed Denial of Service (DDoS) attacks. This paper presents a novel detection framework that employs frequency-domain analysis to uncover hidden attack patterns within PacketIn message fluctuations. To further refine detection accuracy, eXplainable AI (XAI) is integrated to optimize the detection accuracy of unseen types of DDoS attacks and enhance the interpretability of the model. Our approach enables a more precise attack classification while minimizing false positives using XAI-driven knowledge transfer. Experimental evaluations confirm that this method significantly strengthens SDN resilience against evolving DDoS threats, providing a more adaptive and intelligent defense mechanism.