Federated reinforcement learning (FRL) for unmanned aerial vehicles (UAVs) in mobile edge computing (MEC) typically relies on fixed-strength proximal constraints or Kullback–Leibler (KL) divergence penalties in distribution space, both of which break down when UAVs in a fleet have different rotor disc areas, coverage radii, and maximum accelerations: fixed coefficients cannot adapt across training phases, and KL estimates become noisy when heterogeneous UAVs explore different state regions. FedDPR (Federated Dual Proximal Regularization) replaces the KL penalty with dual L2-norm proximal terms in parameter space for state-independent gradients, and anneals the global coefficient over FL rounds to transition from fleet-wide knowledge sharing to per-UAV hardware specialization. FedDPR outperforms FedProx, FMARL, and FedKL in energy efficiency, average waiting time, and cumulative reward.
The 5G low-density parity-check (LDPC) standard employs a fixed 2-block puncturing strategy, where the first two information blocks are punctured to transmit additional parity blocks. This approach assumes that the extension gain outweighs the puncturing loss. However, this study reveals that this assumption is not universally valid, particularly in low-iteration or low-code-rate regimes. To address this, we investigate a flexible puncturing framework incorporating 0-block and 1-block schemes. Simulation results demonstrate significant performance gains in iteration-constrained or low-rate scenarios. Finally, we develop an operational optimization map to select the most effective puncturing strategy based on the operation environment.
The integration of AIoT in healthcare allows continuous monitoring of patients and real-time clinical decision-making. Energy inefficiency of batteries in edge devices, as well as latency and accuracy limitations, are among the key concerns in the domain. In this study, a Reinforcement Learning (RL) driven Variational Quantum Algorithm (VQA)-based multi-objective approach (RL-VQA) to optimize AIoT for energy efficiency by incorporating quantum circuit parameter optimization is introduced. In this context, the intelligent optimization of the system-level design is attained through dynamic adjustment of variational parameters using reinforcement learning while considering energy efficiency, latency, and diagnostic accuracy. The problem setting is designed as a constrained optimization problem, allowing efficient search for possible solutions in the complex problem space through hybrid quantum-computational models. Results show that the approach yields up to 31.7% reduction in energy, 24 ms improvements in latency, and 1.1% higher accuracy in comparison to classical and static quantum approaches.
Open Radio Access Network (O-RAN) disaggregates the RAN into standardized units coordinated by Near-Real-Time (Near-RT) and Non-Real-Time (Non-RT) RAN Intelligent Controllers (RICs), enabling closed-loop optimization. Proactive control requires accurate prediction of traffic demand, user equipment (UE) mobility, and resource consumption before congestion or service level agreement violations occur. Yet machine learning (ML)-based proposals remain fragmented across granularities, interface placements, and evaluation environments, with no unified framework for comparing designs or assessing evidence quality.This paper surveys recent works on ML-based traffic and UE behavior prediction that explicitly feed O-RAN control loops. We classify each work along four axes: (i) ML tier, separating architectures with structural inductive bias from standard deep learning and classical methods; (ii) traffic target at UE, slice, and cell granularities; (iii) evaluation environment across live deployments, emulation, and simulation; and (iv) O-RAN interface placement across A1, E2, O1, and O2.Our analysis reveals four headline findings. First, a forecast-then-act pipeline, composing a predictor with a downstream deep reinforcement learning or optimization controller, dominates closed-loop design across granularities. Second, standard deep learning predominates, whereas architectures with structural inductive bias remain a minority. Third, evaluation is dominated by custom, unstandardized simulators, with no surveyed work performing live closed-loop actuation on an operational deployment. Fourth, open-source reproducibility is near-absent, with only one complete release of code, real data, and evaluation scripts. We then identify open challenges in per-UE forecasting, standardized data exposure, and open-source evaluation infrastructure, and outline a research roadmap toward 6G predictive intelligence in O-RAN.
This paper proposes an Adversarial-Aware Multi-Agent Deep Deterministic Policy Gradient (AA-MADDPG) framework integrating adversarial risk analysis with deep reinforcement learning for robust resource allocation and offloading in heterogeneous aerial access networks. We model adversarial behavior using three rationality paradigms, including Nash equilibrium, level-k thinking, and prospect maximizing, employing Bayesian model averaging for robust adversarial action prediction. The framework enables collaborative decision-making among aerial access tiers and IoT devices while maintaining attack resilience. Simulations demonstrate that AA-MADDPG reduces task drop rates by 27% and energy consumption by 19% compared to baselines.
Immersive services such as virtual reality (VR) and holography demand multi-gigabit throughput and ultra-low latency. To meet these demands, THz-band directional communication is essential but vulnerable to LoS blockages in dynamic responsive spaces. To address this issue, we propose a Responsive-Space Directional MAC (RS-DMAC) based on Dueling Double DQN. RS-DMAC jointly learns the selection of beams and relay nodes using an observation-driven RL formulation based on communication-level measurements. Simulation results demonstrate that RS-DMAC achieves higher throughput, lower latency, and improved link reliability compared to conventional schemes for immersive service requirements, while maintaining a reasonable energy-efficiency trade-off.
Energy efficiency has emerged as a critical design objective for sixth-generation (6G) radio access networks (RAN) due to their significant environmental and economic impact. According to multiple industry and academic projections, 6G networks could consume two to three times more energy than fifth-generation (5G) networks, particularly since the RAN accounts for approximately 73% of total network energy consumption (within a commonly reported range of 70%–80%). This survey provides a comprehensive overview of energy-efficient mechanisms for 6G RAN across four key optimization domains: time-domain techniques (e.g., discontinuous reception/transmission and adaptive synchronization signaling), frequency-domain methods (e.g., bandwidth part operation and carrier activation), spatial-domain approaches (e.g., dynamic antenna adaptation and beamforming), and power-domain optimizations (e.g., power amplifier efficiency and dynamic power control). Beyond conventional mechanisms, the survey reviews emerging artificial intelligence (AI)/machine learning (ML)-driven optimization approaches, including traffic prediction, reinforcement learning, and federated learning frameworks, highlighting how 6G networks can achieve AI-native energy awareness from the design stage. As third generation partnership project (3GPP) standardization evolves from Release 15 through Release 21, with particular focus on the energy-native directions of Releases 19–21, energy-saving technologies are increasingly aligned with quantitative and measurable energy-efficiency requirements for 6G systems. Additionally, the survey examines several emerging enabling technologies expected to influence the energy efficiency of future networks, including reconfigurable intelligent surfaces (RIS), terahertz communications, and holographic multiple-input and multiple-output (MIMO). It also identifies several critical standardization challenges, such as AI/ML framework interoperability, the lack of consensus on energy measurement methodologies, and ongoing debates regarding unified energy-efficiency metrics. Our findings suggest that achieving the targeted 50% reduction in overall energy consumption for 6G will require coordinated advancements across multiple dimensions, including hardware efficiency, architectural design, and intelligent control mechanisms. These developments are expected to align with the 3GPP Release 21 specifications anticipated for 2027–2028.
Face recognition methods are primarily designed for single-image analysis, even though video-based recognition has seen a dramatic increase in popularity in edge security and surveillance applications. Typically, a video template is constructed from the features of individual frames. Feature norms are commonly used as weights in the construction process, as they correlate well with the usefulness of samples for recognition. Classical training approaches directly optimize only the angular distances, in turn also guiding the feature norms. This can lead to suboptimal alignment between feature norms and the usefulness (utility) of samples, resulting in subpar video performance. Motivated by this insight, we propose the UVFace methodology, which presents an extended feature norm alignment branch. Through careful design of the quality ranking step, which produces feature norm labels and a new feature norm loss, UVFace improves performance over the reproduced AdaFace baseline on video-oriented benchmarks while retaining strong image-based performance. Code is available at https://github.com/LSIbabnikz/UVFace.
Small objects have become increasingly common in civil, industrial, and security sensitive environments, driving a strong demand for reliable vision-based detection and segmentation. While deep learning (DL) models have achieved impressive performance, their practicality is limited by the small sizes of objects, cluttered backgrounds, motion blur, domain shift, and the growing cost of training and inference. Quantum machine learning (QML) has recently emerged as a potential research direction for exploring alternative feature representations, optimization strategies, and learning paradigms under resource constrained conditions. However, its practical efficiency benefits for object detection and segmentation remain largely preliminary and context dependent. The literature on quantum methods for detection and segmentation is scattered across the QML, aerial vision, and remote sensing domains. We provide a structured review of quantum, hybrid quantum–classical (HQC), and quantum approaches for detection and segmentation. We classify existing methods based on four aspects: how features are encoded and represented in quantum systems, the use of quantum kernels, variational quantum circuits (VQC) for classification and detection, and quantum architectures combined with modern detection and segmentation models. We summarize the commonly used datasets for quantum techniques, evaluation protocols, and computational constraints, comparing the reported performance trends, scalability levels, and practical limitations of quantum-based methods. Finally, we identify the open challenges, including data-efficient learning, robustness to domain shift, and end-to-end quantum segmentation, along with the research directions relevant to bridging the gap between QML and real-world object perception systems.
License plate recognition (LPR) is an important component of intelligent transportation systems. With the rapid development of deep learning, LPR has evolved from handcrafted pipelines toward data-driven detection and recognition models with substantially improved accuracy and efficiency. However, the progress reported in existing studies is often difficult to interpret clearly due to considerable differences in dataset splits, evaluation protocols, preprocessing strategies, task definitions, and hardware settings. This review examines LPR from a system-level perspective. We discuss dataset bias, limitations of benchmarks, different evaluation criteria, and the growing importance of restoration-aware LPR, including super-resolution for degraded license plate images.
Ransomware remains a significant cyber threat for legacy and recent systems. This study revisits behavior-based detection and introduces RansomFisher, a novel decoyless ransomware detection system that monitors file access patterns. RansomFisher defines four features to characterize process behaviors and distinguish ransomware from benign applications: write ratio, bulk ratio, write count ratio, and continuous write count. RansomFisher was evaluated on eight recent ransomware samples and 96 benign processes, including four applications that exhibit ransomware-like behavior and 92 system services. RansomFisher detected ransomware with an average file loss of only 15.3 files before detection and a negligible average overhead of 3.7%.
Video-based object detection systems often incur high computational and energy costs in real-time applications. This study proposes a hybrid frame skipping strategy for energy-efficient vehicle detection using YOLOv8n. The method combines motion-aware conditional inference, periodic detector refresh, and online threshold calibration. Experiments on 40 UA-DETRAC videos show that the proposed approach preserves detection accuracy while reducing computational workload. The best configuration (k=3, p=50) achieves only a negligible F1 reduction (ΔF1=-0.0008), reduces mean GPU energy consumption by 34.3%, and increases throughput from 58.7 FPS to 88.7 FPS. Additional analyses confirm favorable accuracy–efficiency trade-offs.
Modern low Earth orbit (LEO) satellite networks use dense inter-satellite links (ISLs) and multipath connectivity to support low-latency communication. This paper proposes utilization-aware multipath traffic splitting over delay-feasible paths in LEO satellite networks. We construct delay-feasible candidate paths using a K-shortest path procedure and apply an L2 traffic splitting method that minimizes aggregate squared link utilization. Compared with Equal Split and LP Split, the proposed L2 splitting method better balances traffic across the network. Simulations on a 72-node LEO topology show that L2 splitting with K=3 reduces average delay from 74.85 ms to 62.40 ms and improves delivery ratio from 72.6% to 82.8% compared with single-path routing. The robustness of the proposed method is further evaluated under realistic traffic scenarios, against adaptive routing baselines, under warm-started re-optimization, and at larger constellation scales. These results suggest that L2-based traffic splitting is a practical approach for software-defined traffic engineering in future LEO satellite networks.
False data injection attacks (FDIAs) represent one of the most significant cyber–physical threats to modern smart grids, as they can corrupt measurement data used in state estimation while evading conventional residual-based detection. This paper provides a structured review of FDIA modeling and detection methodologies for cyber–physical power systems, covering both conceptual and mathematical formulations with attention to adversarial knowledge, measurement topology, and stealth bypass conditions. In this review Detection methods are organized into four families: statistical models, classical machine learning, feedforward deep learning, and recurrent neural networks. A comparative analysis evaluates their detection accuracy, computational cost, temporal modeling capacity, and real-time feasibility. Results show that architectures capturing temporal dependencies particularly LSTM and GRU consistently outperform statistical baselines and static classifiers. GRU offers a favorable trade-off between model complexity and performance. The paper further identifies concrete research gaps, including multi-attack detection, cross-topology generalization, physics-informed learning, explainable AI, privacy-preserving training, and quantum-enhanced approaches.
Terahertz (THz) V2X networks with integrated sensing and communication (ISAC) face severe blockage and energy constraints in urban environments. This letter studies relay-assisted communication with vehicle-side sensing QoS constraints. We formulate a joint beamforming optimization problem to minimize power consumption under strict communication and sensing constraints. An efficient iterative algorithm utilizing Matrix-Extended Generalized Lagrangian Dual Transformation (MEGLDT) and Block Coordinate Descent (BCD) is developed. Simulations demonstrate that the proposed scheme yields significant power savings compared to benchmarks, particularly in high-density scenarios. These findings validate the framework’s scalability and efficacy for next-generation V2X applications.
Existing crack segmentation methods perform well only on the specific datasets they were trained on, limiting their use in real structural health monitoring where cracks vary widely in scale and appearance. This paper proposes ASCBR-Net, a U-Net based architecture that embeds SE and spatial attention within each residual block and places a multi-rate ASPP at the bottleneck to capture multiscale context before spatial reconstruction. Evaluated on four merged public datasets under 5-fold cross-validation, ASCBR-Net outperformed seven baseline methods with statistically significant gains (p<0.01), improving F1-score by up to 9.30% and IoU by up to 7.61%.
Hierarchical federated learning (HFL) reduces cloud-facing traffic by aggregating client updates at intermediate edge servers, but plaintext aggregation exposes model updates to honest-but-curious infrastructure. This paper studies a client–edge–cloud HFL architecture protected by the Cheon–Kim–Kim–Song (CKKS) approximate homomorphic-encryption scheme and evaluated under wireless impairments. Clients encrypt packed model updates, edge servers aggregate ciphertexts, and the cloud completes the second aggregation stage without accessing plaintext updates. A trusted key holder alone decrypts the global update. The wireless model accounts for heterogeneous bandwidth, fading, packet loss, retransmissions, and deadline-based client admission. Experiments on Fashion-MNIST use 20 clients, four edge servers, Dirichlet non-IID partitioning with concentration α=0.5, and 100 global rounds. Four variants are compared: plaintext HFL, CKKS-HFL, wireless plaintext HFL, and wireless CKKS-HFL. Under reliable links, plaintext and CKKS curves nearly overlap and exceed 91% final accuracy, while CKKS relative aggregation error remains below approximately 6×10−10 in the ideal-link case. Wireless plaintext HFL remains close to the reliable-link baseline because almost all clients are received each round. In contrast, wireless CKKS-HFL admits only about one to seven clients per round because encrypted updates are much larger, causing noisier convergence and lower final accuracy of approximately 89%. HFL nevertheless reduces encrypted edge-to-cloud traffic from roughly 670 MB at the client tier to about 470 MB in the ideal CKKS case, with larger reductions when wireless admission limits participation. These results show that CKKS numerical error is negligible, whereas ciphertext-induced communication pressure is the dominant practical bottleneck.
This survey presents a system-level architectural analysis of quantum computing for portfolio optimization in FinTech, focusing on problems that can become NP-hard under realistic discrete or combinatorial constraints and that face scalability challenges in classical cloud infrastructures. Based on a systematic review of 30 representative papers (2020-2025), three dominant paradigms — Quantum Annealing (QA), Variational Quantum Algorithms (VQAs), and Quantum Machine Learning (QML) — are evaluated specifically by focusing on their constraint enforcement strategies and latency profiles. The analysis suggests that QA and hybrid CQM approaches currently demonstrate significant near-term operational readiness for constraint-heavy portfolios due to its native solvers and lower communication overhead compared to iterative gate-based methods. Conversely, while VQAs offer high expressivity via mixers, they are operationally limited by the communication complexity of hybrid loops over public networks. Furthermore, QML approaches are shown to shift the bottleneck from optimization to state preparation, where data encoding latency often negates inference advantages. These engineering insights are synthesized into a practical implementation roadmap, advocating for distributed hybrid architectures to bridge the gap between theoretical quantum advantage and low-latency financial deployment.
Secure Real-time Transport Protocol (SRTP) provides end-to-end media confidentiality in Real-Time Communication (RTC) systems. However, since SRTP embeds application-specific semantics into its unencrypted headers and encrypts only the payload, this metadata remains observable—potentially enabling traffic analysis by passive adversaries even under encryption. We investigate whether plaintext SRTP header extensions can enable packet-level service classification in modern RTC deployments. We analyze six popular RTC applications and propose an ML-based classification pipeline that utilizes RTP header extension metadata to identify apps without any payload decryption, achieving a robust F1-macro score of 0.95. Furthermore, we propose and evaluate a lightweight, permutation-based countermeasure.
The growing complexity of 5G/6G, IoT, and edge networks exposes the limits of single-agent LLM control in scalability, real-time performance, and reliability. Multi-Agent Systems (MAS) combined with LLMs offer a path forward. This paper surveys state-of-the-art LLM-based MAS for network management. We propose a taxonomy that re-interprets the Centralized/Decentralized/Hybrid split under two networking-specific axes (a strict LLM-core inclusion criterion and an anchoring on the NetOps lifecycle), complemented by a role-based lens (Coordinator, Translator, Negotiator). We finally identify key challenges such as hallucination and inference latency, and outline directions towards trustworthy autonomous networks.
