
Backup resource allocation is essential for service continuity in cloud infrastructures, where hardware failures may disrupt services. Probabilistic protection reduces reserved backup resources by allowing a bounded backup failure probability, but evaluating this probability under simultaneous physical machine (PM) failures and dimensioning capacity with low resource usage remain computationally challenging. We propose probabilistic backup resource allocation models for cloud servers under simultaneous multiple PM failures, with separate formulations for bare metal and virtual machine (VM) deployments. In the bare metal case, the backup failure probability is computed from a binomial distribution. In the VM case, failed VMs are packed by first fit decreasing within a binomial failure model. For VM capacity dimensioning, we develop a logarithmic backoff descent strategy (LBDS) to find feasible low-capacity allocations under the survivability constraint. For large-scale VM instances with repeated VM-capacity patterns, we decompose the problem into typed domains to reduce the evaluated state space and compare the resulting capacity and runtime with a modified baseline adapted to typed-domain large-scale instances. Numerical results show that LBDS matches exact enumeration in seven of eight tested small VM settings and has a worst observed capacity gap of 100 capacity units, while requiring fewer probability evaluations; in the tested typed-domain large-scale setting, the observed required capacity is up to 15.3% lower and the runtime is at least 41.2% lower than that of the modified large-scale baseline.
The next generation of mobile networks is moving artificial intelligence out of a sporadic implementation model and into the managed service paradigm, which requires careful placement, scaling, validation and security of the radio, edge and cloud space. Although federated learning is vulnerable to meet strict privacy requirements, its training process has the properties of a functional service chain, which includes transient participation and intermittent client interactions, competitive service slices, and the need to implement trust and comply with latency and reliability requirements. Secure federated intelligence is described as an orchestrated 6G-optimized service graph in this research, which includes: (i) slice-aware admission and resource assignment; (ii) trust-gated robust aggregation; and (iii) post-quantum secure update exchange as composable tasks. An orchestration workflow with a feasibility-first constraint is planned to maintain constraints at the service level and also ensure the convergence in case of heterogeneous participation and adversity of poisoning. Large scale simulation of three different 6G slice formats illustrate a steady decrease in tail latency and SLA breaches, an improved resistance to malicious clients, and a relatively low control overhead, thus supporting the feasibility of the federated modelling to federated managed service transition.
The rapid evolution of cellular networks has made them a critical infrastructure for modern society, but it has also led to quickly increasing energy consumption. Open Radio Access Network (O-RAN), with its disaggregated architecture, offers significant potential for improving energy efficiency (EE) in next-generation networks. This survey presents a comprehensive review of state-of-the-art O-RAN energy efficiency research, structured around three key dimensions: energy measurement, energy modeling, and energy optimization. We examine both software- and hardware-based measurement techniques, highlighting their strengths, limitations, and applicability on different network functions. The survey then categorizes existing energy consumption models into theoretical and empirical approaches, with a discussion of their methodologies and constraints. Finally, we explore energy optimization strategies defined by O-RAN specifications and beyond, revealing diverse technical directions in academia and industry. The survey also highlights emerging research challenges, including the lack of high-fidelity and generic datasets, the need for standardized evaluation frameworks and methodologies, and the impact of continuous O-RAN evolution on the validity of existing studies. Overall, this work aims to serve as a foundational reference for researchers and engineers to build sustainable and energy-aware O-RAN networks.
Federated Learning (FL) is a promising privacy-preserving machine learning paradigm that facilitates shared model training over distributed devices. Conventional FL employs a synchronous protocol, where each round’s training time depends on the slowest client, leaving faster clients idle after submitting updates, which considerably decreases FL training efficiency. To address this straggler issue, Asynchronous Federated Learning (AFL) has been proposed, where the server asynchronously aggregates received model updates without waiting for stragglers. But it introduces stale updates that can impair model convergence and accuracy. Existing AFL methods mainly address this issue by reducing the influence of stale updates. However, this mitigation strategy, together with clients’ unbalanced reporting frequencies due to varying capacities, biases the global model toward faster clients, which may discourage slower clients with valuable data from participating. To address this issue, we propose a novel Shapley value-based fair AFL method, AFedSV, that promotes fair (uniform) model performance across clients without compromising the overall model accuracy. Specifically, we leverage Shapley value to identify beneficial updates independent of their staleness and emphasize the contribution from stale ones in aggregation to balance the representation of slow and fast clients. Experimental results show that AFedSV outperforms baselines in terms of fairness and accuracy.
The increasing computational demands of Internet of Things (IoT) devices require efficient Mobile Edge Computing (MEC) strategies, especially in remote and underserved areas. Traditional terrestrial MEC solutions often encounter challenges such as computational overload and latency issues. This paper introduces a novel intelligent Deep Reinforcement Learning (DRL)-based approach, named Software-Defined Soft Actor-Critic (SD-SAC), specifically tailored for hybrid Low Earth Orbit (LEO) satellite-edge networks. The proposed framework integrates hierarchical Software-Defined Networking (SDN) with DRL to dynamically optimize atomic task offloading decisions. Unlike conventional methods, our solution categorizes satellites into computational edge nodes and non-computational relay nodes, allowing for optimized resource utilization and enhanced task processing. The SD-SAC algorithm employs real-time network state monitoring and adaptive decision-making to effectively balance computational loads, reduce latency, and maximize the task success rate. Comprehensive simulations validate that the SD-SAC framework significantly outperforms traditional offloading methods by improving Quality of Service (QoS) satisfaction, minimizing delays, and ensuring efficient CPU and bandwidth utilization. This research offers a robust solution for managing computational tasks in dynamic satellite-edge computing environments, demonstrating clear advantages in QoS-aware task atomic offloading strategies.
Flying Ad-Hoc Network (FANET) consists of unmanned aerial vehicles (UAVs) that cooperatively establish wireless communication networks. It has been widely used in communications, transportation, forest fire monitoring and other fields. Due to the highly dynamic topology and boundary disconnection problems in three-dimensional (3-D) environments, designing an efficient routing protocol for FANETs remains challenging. To address these issues, this paper proposes a Boundary-Aware Routing Protocol utilizing Game Theory (BARP-GT). Firstly, BARP-GT defines a 3-D forwarding area by considering network boundary effects. Then, it estimates node degree in different boundary scenarios without relying on Hello message broadcasting, which reduces routing overhead and energy consumption. In addition, the forwarding process of route request packets is modeled as a static game to derive an adaptive forwarding probability and suppress broadcast storms. Simulation results based on Network Simulator-2 (NS-2) demonstrate that BARP-GT achieves better performance than iPipe, QEHLR, and BLPR in terms of end-to-end delay, packet delivery ratio, packet drop ratio, and routing load.
In the ever-evolving Industry 4.0 landscape, logs and events play a critical role in maintaining the reliability and security of complex systems. This is especially true for 5G networks, where lightweight protocols such as CoAP and LwM2M produce massive volumes of structured but often unlabeled log data. Although machine learning and deep learning methods have become common tools for anomaly detection in such environments, they are frequently hindered by practical limitations: scarce labeled data, high-dimensionality, long log sequences, and the lack of realistic attack scenarios in available datasets. To address the above challenges, we adopt a deep clustering framework that learns low-dimensional structured latent representations without requiring labels. TF–IDF statistical analysis excels at detecting repetitive attack patterns and anomalous term frequencies characteristic of threats like distributed denial-of-service attacks and message flooding, while enabling rapid compression and interpretation of large-scale log data. However, TF–IDF alone fails to capture semantic context—for instance, distinguishing between benign timeouts and attack-induced timeout cascades. Conversely, while pre-trained models like sentence-transformer capture rich semantics, their latent spaces are optimized for classification and are often unsuitable for clustering, due to their entangled and non-topological structure. In this paper, we propose a novel hybrid architecture that fuses a TF–IDF-based autoencoder with a sentence-transformer encoder through a cross-attention mechanism. This combination leverages TF–IDF’s statistical sensitivity to repetitive attack signatures while enriching it with semantic understanding, allowing the latent representation to selectively incorporate both statistical anomalies and contextual semantic signals, thereby preserving the interpretability and clusterability of the learned space. We implement this architecture through a Wazuh-based SIEM deployment at the 5G network edge, demonstrating that unsupervised hybrid clustering can deliver effective CoAP anomaly detection in industrial IoT environments without labeled data, intrusive agents, or cloud-dependent processing addressing critical gaps in operational security for private 5G networks.
Ultra-Reliable Low-Latency Communications (URLLC) imposes stringent latency and reliability requirements that are difficult to sustain in dynamic 5G network slicing environments, especially under demand uncertainty and abrupt traffic surges. This paper proposes PRISM-URLLC (Predictive and Reserve-aware Slice Management for URLLC), a prediction-integrated and reserve-aware online slice orchestration framework for URLLC-oriented resource allocation under dynamic multi-slice conditions. The proposed method combines observed demand, short-horizon predicted demand, QoS sensitivity, slice priority, and urgency-related signals within a lightweight score-based allocation mechanism. To improve responsiveness under sudden load escalation, the framework further introduces an uncertainty-aware adaptive emergency reservation policy that adjusts the reserved resource budget according to urgency intensity, predicted URLLC demand, and prediction uncertainty. Rather than relying on a specific forecasting architecture, PRISM-URLLC is designed to incorporate predictive inputs in a modular manner while preserving low computational overhead. The framework is evaluated in an abstract slice-level 5G simulation environment under nominal, high-load, and emergency-surge traffic regimes, and is compared against demand-aware, priority-based, fairness-oriented, and hybrid non-predictive baselines under identical assumptions. Results show that the proposed method improves URLLC-oriented service indicators, including demand satisfaction and latency-related performance proxies, particularly under high-load and emergency-surge conditions, while maintaining competitive fairness, resource utilization, and runtime efficiency. Additional analyses on scalability, robustness to prediction noise, and component ablation further support the relevance of the proposed framework as a practical online orchestration mechanism for dynamic URLLC-oriented slice management.
Over-the-air federated learning (OTA-FL) enables distributed edge devices to collaboratively train a global model without sharing raw data, leveraging the wireless channel's superposition property for efficient aggregation. However, practical deployments face challenges due to limited bandwidth, energy, and computational resources, as well as wireless channel impairments. This paper presents COMPACT-FD, a distillation-enhanced OTA-FL framework that combines knowledge distillation (KD), post-training magnitude-based pruning, and over-the-air (OTA) aggregation to characterize SNR-aware compression regimes. The approach combines KD for accuracy and stability, pruning for model compression, and OTA aggregation for communication and resource efficiency. We analyze the interplay between compression and wireless channel conditions, deriving SNRaware operating recommendations to balance accuracy, model size, and resource constraints. Experiments under varying SNR conditions show that COMPACT-FD achieves substantial reductions in communication and computation costs while maintaining robust accuracy and convergence. Our results provide deployment guidelines for edge intelligence in realistic wireless environments, highlighting the benefits of combining KD, pruning, and OTA aggregation in OTA-FL.