Channel estimation in massive MIMO (Multiple-Input Multiple-Output) systems with one-bit ADCs (Analog-to-Digital Converters) faces severe quantization distortion that degrades conventional estimators and limits downstream detection and precoding. While cGAN (Conditional GAN)-based learning approaches can recover CSI (Channel State Information) from heavily quantized pilots, their susceptibility to adversarial perturbations has received limited attention despite the security sensitivity of wireless physical-layer processing. This paper presents the first framework that combines auxiliary classifier supervision with adversarial training for robust one-bit massive MIMO channel estimation. The proposed AC-GAN (Auxiliary Classifier GAN) discriminator is trained to jointly perform source discrimination (real/fake) and SNR (Signal-to-Noise Ratio) regime classification, encouraging disentangled, regime-aware feature learning that standard cGAN designs cannot achieve. We further integrate PGD (Projected Gradient Descent)-based adversarial training with curriculum scheduling into a minmax learning objective to improve robustness under white-box attacks. Simulations on a $\mathbf{6 4}$-antenna, 8-user uplink demonstrate that AC-GAN attains 15.3% lower NMSE (Normalized Mean Squared Error) than cGAN under clean conditions (-23.3 dB vs. $-20.8 \mathbf{~ d B}$) and achieves $\mathbf{7 8. 6 \%}$ robust accuracy under strong PGD attacks $(\varepsilon=0.1)$, outperforming adversarially trained cGAN $(37.4 \%)$ by 41.2 percentage points. The results show consistent gains across SNR regimes, pilot lengths, and attack types, including unseen attack strategies, supporting the use of auxiliary supervision as a practical and effective approach to more secure DL (Deep Learning)-enabled channel estimation.
Noise in magnetic resonance imaging (MRI) degrades image quality and adversely affects both clinical interpretation and subsequent signal processing based analysis. This paper presents a comparative evaluation of classical image denoising filters and a learning-based approach for MRI noise reduction. Specifically, Gaussian, median, and bilateral filtering techniques are analyzed alongside a conditional generative adversarial network (cGAN). Experiments are conducted on the Brain Tumor Segmentation dataset, comprising 3,064 T1-weighted contrast-enhanced MRI slices from 233 patients. Performance is quantitatively assessed using standard signal processing metrics, including peak signal-to-noise ratio (PSNR), structural similarity index (SSIM), mean squared error (MSE), mean absolute error (MAE), entropy, and sharpness. Among the classical methods, bilateral filtering achieves the highest PSNR (36.36 dB), the lowest MSE (0.000231), and the highest sharpness value (0.1431), while Gaussian filtering attains the best SSIM score (0.9514). The cGAN-based approach demonstrates competitive performance, yielding a PSNR of approximately 33.0 dB, an SSIM of 0.92, the lowest MAE (0.0130), and the highest entropy (13.99), indicating improved perceptual quality and effective noise suppression. These results highlight that while classical signal processing filters particularly bilateral filtering provide strong numerical performance, learning-based denoising offers complementary advantages in perceptual fidelity and detail preservation, supporting advanced MRI preprocessing pipelines.
Mixed Reality (MR) applications rely on continuous camera capture to seamlessly integrate virtual and real-world content, raising significant privacy concerns due to the inadvertent exposure of sensitive visual information. Such information can be leaked not only through the direct visual appearance of private objects, but also through indirect visual cues. Addressing these cues is therefore as critical as direct object obfuscation, as side channels such as shadows and reflections may still reveal sensitive information about the object. Consequently, ignoring these side channels can result in inadvertent privacy leakage in a real-world MR scenario. Although several privacypreserving frameworks have been proposed, existing approaches focus on a single form of visual leakage, limiting their effectiveness in real-world MR scenarios. In this paper, we propose a lightweight privacy-preserving framework designed for real-time mobile MR deployment that jointly addresses shadow and reflection obfuscation within a unified pipeline. Experimental results demonstrate improved end-to-end performance compared to the existing frameworks, reducing the inference time of the reflection and shadow pipelines by $4.1 \times$ and $2 \times$, respectively.
Software-Defined Networking (SDN) improves programmability via centralized control, but exposes availability risks under flooding attacks that degrade controller-mediated operation. While most SDN security work focuses on detection, effective prevention requires timely, network-wide mitigation, and multi-domain deployments often cannot pool raw traffic for centralized learning. This paper evaluates a federated reinforcement learning intrusion prevention system (FRL-IPS) that learns rate-limiting mitigation policies while keeping traffic local. We implement a multi-controller SDN emulation with three domains and train a Double Deep Q-Network (DDQN) agent, coordinated via Flower-based federated learning, under homogeneous (IID-like) and heterogeneous (non-IID) domain conditions. Results show that federated training can reach prevention behavior comparable to centralized training at steady state, but requires more training to stabilize, with heterogeneity further increasing transient variability and stabilization time. These findings support the feasibility of privacy-preserving, prevention-oriented learning in SDN, while highlighting the responsiveness cost introduced by federation and non-IID conditions.
Accurate and energy-efficient localization remains a fundamental challenge in large-scale IoT networks, where resource constraints and dynamic signal conditions limit the effectiveness of conventional approaches. Purely stochastic models often fail to capture complex spatial signal patterns, while deep learning models alone typically ignore temporal structure and incur high computational cost. We propose a hybrid localization framework that unifies stochastic temporal modeling with deep neural feature extraction. Specifically, we introduce Markov-CNN, which integrates convolutional neural networks (CNNs) for spatial feature learning with Markov-based temporal state modeling, and Markov-CNN-PCA, which further incorporates principal component analysis (PCA) to reduce feature dimensionality and computational overhead. The architecture enables structured temporal smoothing over learned embeddings while preserving scalability for resource-constrained IoT deployments. We conduct extensive simulations comparing our models against standalone Markov and CNN baselines across localization accuracy, energy consumption, and end-to-end latency. Results demonstrate that hybridization yields consistent and substantial improvements across all metrics. In particular, Markov-CNN-PCA achieves the best performance-efficiency trade-off, delivering superior localization accuracy while reducing energy usage and inference latency, thereby enabling real-time deployment.
Sub-Terahertz (Sub-THz) communication is a key candidate for future wireless systems due to the availability of large bandwidths. However, it suffers from severe propagation loss, and is highly sensitive to hardware and alignment impairments. This paper presents an end-to-end virtual Software-Defined Radio (SDR) testbed operating at 180 GHz, implemented in MATLAB/Simulink. The system integrates QPSK baseband framing, synchronization, and equalization with a hardware-aware RF chain based on an AD9361 intermediate-frequency transceiver and cascaded frequency extension to the sub-THz band. Directional horn antennas and a multipath fading channel are included to model realistic propagation effects. Simulation results validate correct frame recovery and show the impact of antennas misalignment as well as fading on the received spectrum. After digital compensation, the receiver achieves stable constellation recovery with low error vector magnitude and bit-error rate, demonstrating the feasibility of the proposed SDR-based sub-THz system model.
This paper explores the essential function of Identity and Access Management (IAM) in reducing insider threats and enhancing the cybersecurity resilience of organisations. Insider threats, whether deliberate or unintentional, pose significant challenges to data integrity, operational stability, and regulatory compliance. With the growing adoption of cloud-based infrastructure and remote work models, identity and access management has transformed from a static access-control mechanism into a dynamic, intelligence-driven framework. The research employs a qualitative interpretive methodology, integrating contemporary academic and industry literature from 2022 to 2025 to examine how emerging technologies such as Artificial Intelligence (AI), Zero Trust Architecture (ZTA), and blockchain enhance the effectiveness of IAM. The findings indicate that AI-enhanced IAM frameworks significantly improve anomaly detection and privilege management, whereas blockchain provides transparency and immutability for access governance. Moreover, following international standards such as NIST SP 800-207 and ISO/IEC 27001 ensures compliance and facilitates interoperability across sectors. The results highlight that successful IAM implementation involves not only technology but also a blend of human, procedural, and regulatory factors. The paper concludes that future advancements in IAM will rely on integrating explainable AI, decentralised identity models, and quantum-resilient architectures to ensure adaptive, ethical, and secure identity governance within evolving digital ecosystems. As this study employs a qualitative, interpretive methodology based exclusively on secondary sources, the findings are conceptual. Empirical validation through longitudinal case studies and controlled deployments constitutes an important direction for future research.
A task-aware semantic communication framework for 6 G short packet transmission is presented in this paper. The limitations of single-variable power optimisation are addressed through a joint design in which reliability, blocklength, power and compute level are optimised together. A two-stream source model is considered so that critical and auxiliary semantic content are protected differently under finite-blocklength constraints. A task distortion model is adopted in which the semantic distortion is reduced both by the transmitted semantic bits and by the available computing resources. The wireless transmission energy and computation energy are then minimised under a target semantic distortion requirement. Both additive white Gaussian noise (AWGN) and Rayleigh fading settings are examined. The reported results show that the proposed joint method achieves the lowest energy among the comparable two-stream schemes over the whole tested distortion range from $D_{0}=0.12$ to $D_{0}=0.35$. At the representative operating point $D_{0}=0.2186$, the proposed method requires a total energy of 106.47, the equal protection baseline requires 138.79 and the fixed blocklength baseline requires 120.88. These values correspond to energy reductions of 23.3 % and 11.9 %, respectively. It is also shown that the best compute level is usually obtained at a moderate value rather than at the smallest or the largest value. Under Rayleigh fading with outage constraint $\delta=10^{-2}$, the required energy is more than 80 times higher than in the AWGN case. These results demonstrate that semantic short packet transmission should be designed using a joint cross-layer optimisation framework rather than power control alone.
Recent studies quantify post-quantum cryptography (PQC) overhead in isolated TLS handshakes, while intrusion detection research focuses on legacy traffic patterns, leaving the intersection of PQC deployment and security monitoring unexplored. This work bridges this gap with the first systematic assessment of multi-hop PQC networks integrated with AI-based intrusion detection. Through systematic evaluation of four deployment configurations, we quantify that multi-hop PQC incurs $17-18 ~\text{ms}$ latency overhead, exceeding single-hop predictions yet operationally feasible. Critically, Isolation Forest maintains 98.11 % detection accuracy despite PQC-modified traffic, significantly outperforming Random Forest $(94.75 \%, p<0.001)$. These findings establish that networks can adopt quantum-resistant cryptography without compromising threat detection, providing foundations for integrated security architectures. While results are simulation-based, they offer critical insights for guiding realworld PQC-IDS deployments.
Vehicular visible light communication (VLC) can complement V2X, but links are sensitive to LoS/FoV alignment, mobility, and ambient-light noise. In treestructured VLC-VANETs, these effects lead to parentlink degradation and reactive reattachments that increase tail delay and packet loss. We propose a learning-assisted proactive reattachment trigger in which each vehicle runs a lightweight local predictor to anticipate imminent QoS violations from observable features (parent-link SNR, short-term SNR trend, relative motion, and basic topology/QoS indicators). When the predicted risk exceeds a threshold, the vehicle proactively switches to a feasible neighbor using the same QoS/SNR parent-selection score as the baseline, while dwell-time, hysteresis, and cooldown safeguards bound ping-pong and signaling overhead. Simulations across vehicle densities and illumination levels show higher packet delivery and lower high-percentile delay than reactive baselines, with reattachment churn kept bounded.
This study investigates how artificial intelligence is reshaping the requirements and expectations of modern incident response (IR) frameworks. Drawing on survey data from 194 cybersecurity practitioners and a Delphi-informed expert review process, the research identifies critical limitations in legacy IR models, including lifecycle rigidity, inadequate ethical oversight, and difficulty integrating AI-driven indicators. The study proposes a Modular AI-Ready Incident Response (MAIR) framework designed to enhance scalability, autonomy governance, and decision transparency across the Sense-Decide-Act-Learn cycle. Results show strong practitioner consensus for industrywide framework modernization and the need for adaptive, AIintegrated architectures. The paper concludes with a roadmap for validating MAIR through real-world SOC pilots and iterative refinement, positioning it as a forward-looking foundation for next-decade cybersecurity operations.
Context-aware data quality assessment enables the selection of quality checks that are aligned with data usage scenarios. While existing frameworks can model context knowledge and derive context-aware data quality assessment plans, their operationalization into executable validation logic remains largely manual. This paper presents a retrieval-augmented execution framework that bridges this gap by automatically translating validated context-aware assessment plans into executable code. The proposed approach embeds and retrieves dataset context representations, while preserving assessment plans as authoritative knowledge stored in a Neo4j knowledge graph. Retrieved assessment plans are combined with the input dataset context to form a structured augmented prompt that strictly constrains a large language model to act as a translator rather than a reasoning or decision-making component. The model is used solely to generate executable validation logic without introducing new quality dimensions or assumptions. The framework is evaluated using a real radiation wireless sensor dataset collected from a monitoring station. The evaluation demonstrates the feasibility of generating valid and executable Python data quality validation code.
This paper presents two antennas of a bioinspired reconfigurable square-ring slot for sub- 6 GHz 5 G applications, functionally motivated by the dynamic opening and closing behavior of plant stomata. By employing a single PIN diode, the proposed antennas simultaneously enable frequency and polarization reconfiguration, thereby reducing circuit complexity, biasing overhead, and power consumption. The antennas are fabricated on a $\mathbf{1. 6 - m m}$ thick FR4 substrate with a relative permittivity of 4.4 and a loss tangent of 0.02. The first antenna is for frequency-reconfigurable mode, the antenna operates in a triple-band state at $3.6 \text{GHz}, 4.95 \text{GHz}$, and 6.85 GHz when the diode is ON, and in a dual-band state at 3.55 GHz and 6.8 GHz when the diode is OFF. Furthermore, the second antenna is for polarization reconfiguration, which is achieved by switching between linear (ON state) and circular polarization (OFF state), with a minimum axial ratio of 0.62 dB at 3.5 GHz. Owing to their compact geometry, low implementation complexity, and multifunctional reconfigurability, the proposed biomimetic antennas represent an efficient and cost-effective solution for adaptive sub-6 GHz 5G and Internet-of-Things (IoT) wireless systems.
Emergency call centers (112/911) remain isolated from Smart City ecosystems, despite the potential for urban sensors, traffic data, and IoT infrastructure to dramatically improve emergency response. This paper presents an integration architecture that fuses voice-based emergency calls with real-time urban data streams for enhanced dispatch decisions. We propose a bidirectional API framework connecting Next-Generation 112 (NG112) platforms with Smart City data hubs, enabling situational awareness enrichment and dynamic resource allocation. Our architecture incorporates a Fusion Engine that aggregates call transcriptions, caller geolocation, fleet GPS positions, traffic conditions, and environmental sensors into a unified operational picture. A Decision Support module recommends optimal resource deployment based on incident severity, proximity, and predicted travel times. Preliminary evaluation on simulated scenarios demonstrates reduced response times and improved resource utilization compared to traditional dispatch approaches.
End-to-end (E2E) latency is a vital performance indicator for edge computing applications, especially for latencysensitive services like autonomous systems, real-time video analytics, and interactive IoT apps. Precise forecasting of E2E latency, involving sensing, data transmission, computing at edge nodes, and response delivery, is crucial for efficient utilisation of resources, placement of service, and service quality guarantee. This paper proposes an E2E latency prediction (ELaP) for edge computing applications. ELaP employs a hybrid CNN-LSTM deep learning model to predict the E2E latency incurred when transmitting a video frame from the client to the edge server and receiving the corresponding notification. ELaP will consider metrics from the edge computing network, such as client, edge server, network, and application, including CPU and network load conditions. Experimental evaluation demonstrates that ELaP reduced the MAE by approximately 86 % compared with the optimal classical baseline and by about 39.5 % compared to the most efficient standalone deep learning model (LSTM) using the YOLOv8 application.
The rapid digital transformation of modern industrial systems has led to the generation of massive volumes of heterogeneous data from sensors, industrial machines, manufacturing platforms, and enterprise information systems. Traditional machine learning approaches often struggle to effectively exploit such large-scale and complex datasets. Recent advances in foundation models, including large-scale transformer-based architectures, have demonstrated remarkable capabilities in learning generalizable representations from massive multimodal data. These models provide new opportunities for developing intelligent and scalable solutions in Industry 4.0 environments. This paper presents a comprehensive survey of foundation models for industrial intelligence. We review the evolution of artificial intelligence architectures from conventional machine learning to large-scale foundation models and analyze key architectural components that enable scalable industrial AI systems. In addition, we examine emerging training paradigms, including large-scale data curation, self-supervised learning, and distributed training infrastructures designed for industrial datasets. The survey further highlights major industrial applications such as predictive maintenance, automated quality inspection, and intelligent manufacturing systems. Finally, we discuss key challenges related to data availability, domain adaptation, realtime constraints, safety, and energy consumption, and outline promising research directions for future industrial AI systems.
Non-Terrestrial Networks (NTNs) are expected to play a key role in sixth-generation (6G) systems by extending Internet of Things (IoT) connectivity beyond terrestrial coverage. Satellite-based IoT enables large-scale deployments in remote and underserved regions; however, it introduces structural challenges related to large coverage footprints, time-varying propagation conditions, and predominantly ALOHA-based uncoordinated random access under massive device density. This paper presents a structured survey of NTN-IoT access technologies and their associated simulation frameworks, focusing on scalability limitations and modeling constraints, particularly those induced by large-scale random access contention in satellite beams. Cellular and non-cellular solutions, including LoRaWAN Chirp Spread Spectrum (CSS) and Long Range Frequency Hopping Spread Spectrum (LR-FHSS), are comparatively analyzed in terms of access robustness and density-dependent performance trade-offs. Existing simulation tools are further examined with respect to their ability to represent NTN geometry, dynamic propagation behavior, large-scale random access contention, and waveform-specific mechanisms. The study highlights persistent fragmentation in current evaluation approaches and emphasizes the need for unified and reproducible modeling methodologies to enable fair performance comparison of future NTN-IoT systems.
The integration of Artificial Intelligence (AI) in medical imaging has transformed diagnostic practices. However, the opacity of many AI models, particularly deep learning systems, poses significant challenges in clinical settings. Explainable AI (XAI) emerges as a cutting-edge solution, aiming to enhance the interpretability and trustworthiness of AI outputs in healthcare. In this study, we provide a comparative analysis of multiple XAI methods to assess their effectiveness in improving interpretability. Additionally, we propose an explainable framework that integrates YOLO11 with Grad-CAM to both detect kidney stones and provide visual explanations of the model's decision-making process. Quantitative evaluation demonstrated high segmentation performance with a precision of 96.6 % and recall of 97.8 %. Qualitative assessment showed that Grad-CAM effectively highlighted clinically relevant areas, increasing model transparency in real-world diagnosis. The proposed approach bridges the gap between accurate automated kidney stone segmentation and clinical explainability, providing a robust tool for computerassisted diagnosis.
Low dose computed tomography is routinely used to minimize radiation exposure of patients, but it has inherent quantum noise that can obscure details in more complicated organs like the pancreas. This paper compares four modern deep learning architectures based on DnCNN, Attention UNet, REDNet and NAFNet for blocking low dose pancreatic CT images. Specifically, the models were assessed quantitatively in PSNR, SSIM, RMSE and EPI, as well as qualitatively through visual inspections of the reconstructed tissues. The results of these analyses indicated that REDNet and NAFNet performed better in general. Such results show the importance of balancing denoising strength with structural fidelity in medical imaging. Future work will focus on hybrid transformer CNN architectures and unsupervised denoising strategies that provide greater strength and generalization across a number of scanners and anatomical sites. Overall, this study represents a strong comparative baseline for low dose CT restoration and provides an avenue to improve pancreatic imaging diagnostic accuracy and image interpretation.
Accurate and reliable player localisation is essential for performance analysis in water polo, where the effectiveness of vision-based systems is limited by aquatic conditions. This paper proposes an ultra-wideband (UWB) localisation system deployed in a full-size water polo field of play, with the aim of assessing positioning accuracy across the field and system coverage by analyzing missing data. An optimised anchor configuration was designed to ensure full coverage of the field of play and minimal localisation error. A measurement campaign is conducted to create a heat map of the error and a heatmap of the missing data. Based on these maps, we evaluate spatial accuracy and data completeness. The results are a mean positioning error of 16.6 cm, with errors below 20.2 cm in 90% of cases, while maintaining an average missing data rate of approximately 1% across the entire pool. Dynamic experiments further confirmed the system's robustness under motion conditions. Applying a spline-based filter improved trajectory smoothness and reduced large localisation errors reducing the error by 2.2 cm in 90% of cases and a standard deviation of 2.6 cm. However, under these conditions, the average missing data is $12,7 {\%}$ when the tag is moving. In addition, the acquisition frequency is reduced to an average of 9 Hz. Furthermore, analysis indicates that there is no correlation between positional error and percentage of missing data. These results show that the proposed UWB deployment achieves sub- 30 cm accuracy with complete coverage in static and dynamic condition, thus validating its suitability for water polo performance and tactical analysis.
