In the age of digital transformation, the rise of Big Data has fundamentally altered how organizations store, process, and utilize information. This whitepaper provides a comprehensive analysis comparing Big Data with traditional data systems, data warehousing, business intelligence (BI), artificial intelligence (AI), data science, and NoSQL databases. By exploring key differentiators such as volume, variety, velocity, and processing capabilities, this paper aims to shed light on how Big Data has reshaped modern technology infrastructures and its role in advancing analytics, decision-making, and operational efficiency.

High-resolution underwater imagery is essential for advancing marine exploration, offering critical data to support resource analysis and discovery. However, underwater images frequently face issues like color loss and diminished contrast. To overcome these challenges, we propose CCMCS-Net, a network for enhancing underwater images that operates across multiple color spaces and is structured in two stages: color correction and contrast augmentation. A Color Correction Subnet (CC-Net) and a Multi Color-space Stretching Subnet (MCS-Net) are the two primary parts of the network. Initially, the Static Correction Module (SCM) leverages green channel data to adjust the red and blue channels, correcting color distortion in degraded images. At the same time, the Dynamic Correction Module (DCM) obtains weight mappings through end-to-end training. It combines SCM to achieve adaptive dynamic compensation of the image, which generalizes the deviation during correction to a certain extent. Next, the RGB, Lab, and HSI color distributions are individually adjusted by applying histogram stretching to the corrected image. The adjusted channels are then concatenated and fed into a Dynamic Fusion Module (DFM) for feature fusion across color spaces. This further enhances image visibility and contrast. The DCM and DFM are trained on our modified U-shaped network architecture. The network is able to adaptively select and emphasize the most useful features for feature aggregation so that the aggregated features better promote high-quality image reconstruction. Extensive experiments demonstrate that CCMCSNet delivers outstanding results in both visual assessments and objective metrics for real and synthetic underwater images.

Ultra-Reliable Low-Latency Communication (URLLC) is a key enabler for many next-generation wireless network applications, including autonomous vehicles, industrial automation, and telemedicine. It is designed to meet stringent requirements of latency as low as one millisecond and reliability levels exceeding 99.999%. These characteristics make URLLC essential for mission-critical applications where delays or failures could result in severe consequences. This paper provides a comprehensive review of URLLC research, categorizing existing studies based on various URLLC applications. Additionally, we identify key challenges, including resource allocation, interference management, and the integration of emerging technologies like AI and edge computing. Open research problems of future wireless networks are discussed, providing potential directions for advancing URLLC capabilities. This paper serves as a detailed reference for researchers by highlighting state-of-the-art methodologies and emerging directions in URLLC.

This paper presents a novel framework based on edge computing, implemented using Kubernetes orchestration, to optimally offload the computational tasks required for centralized control of multiple robotic agents. Edge-based centralized control architectures are prone to failure due to communication delays. The proposed framework computes the maximum round-trip time delay for which the system remains stable and modifies the controller parameters to ensure the control computation within the critical time. For higher processing and communication delays, the complexity of the controller needs to be reduced by reducing the number of agents, the prediction horizon, and the efficient use of edge resources. The edge resources are dynamic, and the controller needs to be designed to guarantee the online computation within a desired time. A dynamic resource allocation method (based on an approximate function of the controller parameters, complexity, and computational resources) is proposed to design the controller parameters to ensure the bounded computation time. To validate the effectiveness of the proposed approach, we conduct experimental evaluations that analyze system behavior under various conditions, providing valuable insights into the performance, scalability, and robustness of multi-agent control systems deployed on edge infrastructure.

As the global demand for energy transition and transport decarbonization intensifies, the development of advanced magnetizable materials becomes crucial for supporting large-scale applications. This study presents the optimization of MAGMENT composites, which are produced using recycled ferrite aggregates combined with binders such as cement, asphalt, or epoxy. These composites are engineered to achieve high magnetic permeability and low core losses, key characteristics for efficient energy systems. Our results demonstrate that by fine-tuning the aggregate size and volume fraction, permeability can be significantly enhanced, with volume fractions above 65% showing the most promise. Although cement workability imposes a 73% limit, the performance of these composites still surpasses industry benchmarks, notably the KH-HT 60µ from KEDA, by refining the particle size distribution. Adjusting the Nominal Maximum Aggregate Size (NMAS) from 4.5 to 19 mm changes permeability from 40 to 180. The superior magnetic performance of the MC60 ® grade, particularly its minimal core losses, underscores its potential as a leading material in the market. These advancements are vital for applications in wireless charging, both static and dynamic, and in highpower transmission systems, addressing critical needs in sustainable transport and energy infrastructure. The use of recycled materials further aligns with the global push for environmentally responsible technologies.

Human operators in industrial cyber-physical systems (ICPS) (e.g. electric power system) are responsible for critical, real-time decision-making in control centers. They often relying on decision support tools (DSTs) to assist with dynamic decision-making. However, DSTs using multi-modal feedback can sometimes lead to cognitive and information overload, potentially impairing operators' performance, specially during extreme events. This work focuses on human-machine cooperation assisted through advanced tools. The Cyber-Physical Transmission Resiliency Assessment Metric (CP-TRAM) is used as a DST to aid operators by integrating both physical and cyber alerts/alarms, for collaborative decision-making. We developed a laboratory setup for experimental human-in-the-loop validation by replicating a power grid control center environment. Developed CP-TRAM tool with an operator training simulator, displays alarms through the Cyber-Power Alarm Tool. To assess CP-TRAM's impact on cooperative decision-making, we simulate a cyberinduced physical event based on an MITRE ATT&CK Industrial Control System (ICS) framework and evaluate operator responses with and without access to CP-TRAM and the Cyber-Power Alarm Tool. Our study includes 36 participants-18 professional operators and 18 students operator. Results demonstrates improvement in decision making with enhanced speed and accuracy of cyber event detection and response with these tools. Additionally, we incorporate human factors data to further validate CP-TRAM's effectiveness in augmenting operator decision-making.

Cyber-attacks on industrial applications, specifically, phishing and web attacks are the most common data breach vectors and, as such, have attracted significant research attention. To mitigate these types of attacks, many countermeasures based on machine learning (ML) have been proposed. Although MLbased countermeasures are reported to yield satisfactory detection performance on phishing and web attacks, they often require massive amounts of manually labelled email and web request data to build these countermeasures. The manual generation of labels, however, can be laborious, error-prone, and may not be proactive in detecting novel phishing schemes and web attacks since attacks need to be identified and annotated prior to training. To cope with the evolution of web attacks and phishing emails, methods which exploit the vast volumes of unlabelled email and web request data should be adopted. This study therefore proposes self-supervised learning (SSL) based on computer vision (CV) and natural language processing (NLP) techniques to pre-train models on unlabelled email texts and HTTP web requests for the detection of phishing emails and web attacks. By leveraging NLP and CV SSL methods, we pre-train models to learn the structural and contextual representations of unlabelled email texts and HTTP requests. By ensembling the features extracted from the pre-trained models, we obtain robust representations of email text and HTTP request data for the effective detection of phishing emails and web attacks. The experiment results on an imbalanced dataset show that the combined self-supervised pre-trained models outperform other existing works in respect to accuracy, precision, recall and F1-score for phishing email and web attack detection.

Large language models (LLMs) have revolutionized the field of artificial intelligence, achieving unprecedented performance in tasks such as text generation, translation, and reasoning. Despite their capabilities, the enormous size and computational demands of these models limit their accessibility and deployment in resource-constrained settings. Knowledge distillation has emerged as a promising approach to address these challenges by transferring knowledge from a large, complex teacher model to a smaller, more efficient student model. This process retains much of the teacher's performance while significantly reducing the computational and memory requirements. This survey provides a comprehensive overview of knowledge distillation techniques tailored for LLMs. We discuss foundational approaches, such as logit matching and feature alignment, alongside advanced methods that leverage intermediate layer supervision, task-specific adaptations, and multimodal extensions. Applications of knowledge distillation in real-world scenarios are explored, emphasizing its role in enabling efficient deployment of LLMs on edge devices and in low-latency environments. Key challenges are identified, including the preservation of emergent behaviors, domain-specific generalization, and the scalability of distillation techniques for increasingly larger models. The survey also highlights ethical and environmental considerations, such as bias transfer and the carbon footprint of model compression. Finally, we outline future research directions, focusing on adaptive frameworks, integration with other compression techniques, and the development of standardized benchmarks to evaluate distilled models.

Ensuring the cyber resiliency of IEC 61850-based Substation Automation Systems (SAS) is critical for reliable grid operations. This work presents a comprehensive analysis of the performance and security of traditional networking and Software-Defined Networking (SDN) for IEC 61850-based SAS. By leveraging SDN's flow control, proactive traffic engineering, and deny-by-default policies, this study highlights its advantages over traditional networking in critical Operational Technology (OT) applications. Three novel factors for cyber resiliency metrics-network performance, protocol-specific attackability, traffic containment, are proposed, culminating in a unified resiliency score. Experimental evaluations using an enhanced real-time Hardware-in-the-Loop (HIL) testbed, includes GOOSE spoofing, DNP3 Denial-of-Service attacks, and network failovers. Results demonstrates SDN's ability to mitigate impact of cyber threats, enhanced network performance, and reduced attack surface. The findings also indicates the significance of adopting SDN at the process bus for scenarios, if full SDN deployment is not feasible. Additionally, this work provides a metric-based approach to evaluate and compare OT network resiliency across various possible configurations.

This paper presents an advanced hierarchical detection and real-time text recognition system tailored for UAE traffic signs. Our system is designed to support critical applications in autonomous vehicle navigation, assistance for visually impaired individuals, and automated traffic sign maintenance. Emphasizing sensor-based applications our approach uniquely incorporates text recognition within the traffic sign analysis pipeline, addressing the complex challenge of bilingual (Arabic-English) signs-a capability not previously explored in the literature.The system employs a three-stage pipeline, combining object detection, symbol detection, and advanced text recognition, all optimized for high-speed, realworld conditions. Trained on DoTaS, a custom dataset with 1,500 UAE traffic sign images, the system achieves a mean Average Precision (mAP) of 0.9124 using YOLOv8. For English text recognition, PARSeq achieves 89.3% word accuracy, setting a new benchmark for real-time, high-accuracy sign interpretation. Operating at around 250 ms per inference (around 4 FPS), our framework enhances safety, accessibility, and efficiency in urban environments.

In machine learning, dealing with binary imbalanced data classification is challenging due to unequal class sizes, leading to model bias. We propose a unique method that uses filtering, ADASYN oversampling, and ENN cleaning to balance data, improve minority class accuracy, and boost overall model performance, showing significant improvements in AUC, F1, and G-mean metrics.

Diffusion models have emerged as a powerful paradigm in generative modeling, underpinned by robust probabilistic and stochastic principles. These models operate by gradually transforming complex data distributions into simple priors via a forward diffusion process and subsequently learning to reverse this transformation. Despite their success in producing high-quality and diverse outputs, diffusion models face significant computational challenges, particularly during training and inference. This paper explores the theoretical foundations of diffusion models, detailing their formulation, training objectives, and connection to stochastic differential equations. We survey key techniques for enhancing their efficiency, including optimized noise scheduling, architectural innovations, accelerated sampling strategies, and hybrid training approaches. Furthermore, we examine the transformative impact of diffusion models across domains such as image and video synthesis, natural language processing, scientific research, healthcare, and entertainment. While these advancements underscore the potential of diffusion models to drive innovation across industries, challenges remain in computational scalability, domain-specific adaptation, and ethical considerations. We conclude by discussing future research directions aimed at addressing these limitations and unlocking the full potential of diffusion models. This work highlights the evolving landscape of generative modeling and its far-reaching applications, establishing diffusion models as a cornerstone of modern artificial intelligence.

Forecasting resource usage in cloud computing environments is crucial for enhancing operational efficiency and optimizing resource management strategies such as autoscaling and overcommitment. While machine learning (ML) models, especially Long Short-Term Memory (LSTM) networks, have demonstrated high predictive accuracy, this paper questions the necessity of such complex models by revealing an interesting insight: LSTM predictions often resemble shifted versions of the original data. We explore the effectiveness of simpler models that leverage data persistence properties, showing comparable performance to ML approaches. Extensive experiments across multiple cloud datasets validate that resource usage exhibits strong temporal correlation, suggesting that basic shift-based predictions can suffice in many scenarios. This paper advocates for a balanced approach, utilizing ML models only when simpler solutions fall short, thereby promoting cost-effective and practical resource forecasting strategies. The methodology presented in this study focuses on experimental analysis using real-world cloud datasets, highlighting the practicality of persistent forecasting techniques and their applicability in large-scale data centers. Additionally, the paper addresses the scalability and computational efficiency of lightweight forecasting methods compared to complex LSTM networks. By analyzing multiple performance metrics such as mean absolute error (MAE) and average relative delta (ARD), the study underscores the potential of non-ML techniques in reducing forecasting overhead without compromising prediction accuracy.

Computed tomography (CT) is a widely used medical imaging modality which provides invaluable visual representation of various conditions ranging from neurological lesions such as haemorrhage, stroke, tumors etc. to cardiovascular disorders like calcium deposits, pulmonary embolism and many other pathologies. However, the ionizing radiation from the CT machine's x-ray tube has to be kept in check, because overexposure is related to elevated risks for genetic mutation or cancer development. In this work, we attempt to reduce the radiation exposure required for high-quality CT image formation by establishing rank sparsity in principal components' domain and developing a compressed sensing framework based on a novel nonlocal and nonlinear low-rank principal component analysis technique in image denoising, which will be subsequently incorporated as a building block for a sparse-view CT image reconstruction framework under the umbrella of convex analysis. Experiments will show that the proposed strategy provides a viable solution for low-dose CT, outperforming other wellknown nonlocal image restoration models in both denoising and reconstruction tasks. In particular, the proposed method will offer 4 − 10% improvement in root-mean-squared error relative to other nonlocal methods at little extra computational time.

Virtualized RAN (vRAN) matches O-RAN Alliance specifications while transitioning towards virtualized functions on general-purpose computing platforms. However, the energy consumption of these systems remains a major concern. Although this issue has been addressed in the literature, previous works oversimplify routing decisions, overlook the benefits of flexible split choices, or neglect the energy consumption of the transport network. Additionally, most studies employing optimal solutions exhibit very limited scalability due to their high computational time. In this work, we present a comprehensive and efficient Mixed Integer Linear Programming model to minimize the energy consumption of O-RAN systems, addressing the limitations of current approaches. We also design and implement a synthetic data generator to evaluate our model across various network usage profiles, topologies, and reasonable-size networks. We achieved valuable insights and promising results in our evaluation. For example, our results show that when devices require high throughput, the transport network incurs significant energy costs and reduces the centralization rate. We also observed that hierarchical RAN topologies can achieve greater energy efficiency than ring topologies, with our approach enabling up to 15% more centralization while saving around 28% of energy and consuming at least one order of magnitude less time than other strategies.

Machine learning (ML) is transforming industries by automating tasks and generating insights from large datasets. Cloud computing platforms like Amazon Web Services (AWS), Microsoft Azure, and Google Cloud Platform (GCP) provide essential resources that enhance the development and deployment of ML models. This paper provides a comprehensive comparative analysis of these cloud platforms, highlighting their ML services, scalability, cost structures, and integration capabilities. Through case studies and performance benchmarks, we explore how these platforms facilitate different stages of the ML pipeline, addressing the strengths and limitations of each. The study serves as a guide for practitioners and researchers looking to leverage cloud platforms for ML applications [6], [7].

This paper presents the Visual Optical Recognition Telemetry EXtraction (VORTEX) system for extracting and analyzing drone telemetry data from First Person View (FPV) Uncrewed Aerial System (UAS) footage. VORTEX employs MMOCR, a PyTorch-based Optical Character Recognition (OCR) toolbox, to extract telemetry variables from drone Heads Up Display (HUD) recordings, utilizing advanced image preprocessing techniques, including CLAHE enhancement and adaptive thresholding. The study optimizes spatial accuracy and computational efficiency through systematic investigation of temporal sampling rates (1s, 5s, 10s, 15s, 20s) and coordinate processing methods. Results demonstrate that the 5second sampling rate, utilizing 4.07% of available frames, provides the optimal balance with a point retention rate of 64% and mean speed accuracy within 4.2% of the 1-second baseline while reducing computational overhead by 80.5%. Comparative analysis of coordinate processing methods reveals that while UTM Zone 33N projection and Haversine calculations provide consistently similar results (within 0.1% difference), raw WGS84 coordinates underestimate distances by 15-30% and speeds by 20-35%. Altitude measurements showed unexpected resilience to sampling rate variations, with only 2.1% variation across all intervals. This research is the first of its kind, providing quantitative benchmarks for establishing a robust framework for drone telemetry extraction and analysis using open-source tools and spatial libraries.

This article describes numerous classification techniques for detecting numbers in handwritten digits written by different people or with manual input, convolutional neural network (CNN), support vector machine (SVM), K-nearest neighbors (KNN), random forest classifier (RFC). The accuracy of different algorithms such as CNN, SVM, KNN and RFC in the handwriting recognition were compared. In the present work, to build and train neural networks or classifiers, I used the Modified National Institute of Standards and Technology Database (MNIST) dataset that contains 70000 digits with 250 different forms of writing, and the size of each image has 28×28. In this work, I proposed a model to implement a classification algorithm to recognize the handwritten digits. I have compared the results of some of the most used Machine Learning Algorithm like Random Forest Classifier (RFC), K-Nearest Neighbor (KNN), and Support Vector Machine (SVM), and with Deep Learning Algorithms like Convolutional Neural Network (CNN) using Keras with TensorFlow. Using these algorithms, I achieved an accuracy of around 98.76% using CNN, 97.38% using SVM, 96.51% using RFC, and 96.38% using KNN. I found CNN algorithm gives the highest accuracy around 98.76% to detect the handwritten digit. This work paves the way to detect different handwritten digits by different people in different fields.