The Convolutional Neural Network (CNN) models are effective in computer vision strategies and have gained popularity due to their strong performance in visual tasks. Nevertheless, models with architectures such as VGG19 are expensive in terms of computational resources and require huge memory, which limits their usage on low-end devices. The study examines how efficiency can be increased in the model VGG19 by using model compression techniques, like pruning (structured and un-structured) and quantization (8-bit and 4-bit Quantization-Aware Training - QAT). The efficiency of the individual compression approaches was tested by thoroughly exploring the VGG19 with the MNIST, CIFAR-10, and Oxford-IIIT Pet datasets. Each model was evaluated against the baseline based on measures of accuracy, model size, inference time, and complexities of the model, CPU usage, and memory usage. The applied QAT approach reduced the model size by 75% with a drop in computational cost across all methods. In addition, the 8-bit quantitative assessment allowed for substantial system compression alongside increased speed delivery with minimal impact on accuracy. The highest compression and sparsity achieved by 4-bit QAT was 48%, which was not effective as it reduced accuracy on complex datasets, with additional computational overhead on T4 GPU. Structured pruning resulted in faster inference, but unstructured pruning also demonstrated a good result in retaining accuracy and even improving it. To simplify the VGG19 structure, pruning and quantization mechanisms are suggested in order to simplify the architecture to implement the model on edge devices sufficiently, without compromising prediction performance.

Image segmentation is an important aspect of image processing and analysis. Medical imaging segmentation is critical for providing noninvasive information about human body structure that helps physicians analyze body anatomies efficiently. Until recently, various medical imaging segmentation approaches have been presented; however, these approaches are deficient in segmenting abdominal organs due to the significant similarity in their intensity levels. The purpose of this research is to propose a method to facilitate the segmentation of abdominal organs and improve the performance of the segmentation. The core functionality of this research is based on the extraction of rib bone from muscle tissues prior to the application of segmentation. This way, efficient segmentation of abdominal organs can be achieved by isolating the rib bone from the muscle tissues located between the rib bone. The proposed rib bone extraction mechanism is applied to four slices of the MICCAI2007 liver data set to isolate muscle tissues from liver tissues that have significant intensity similarity to liver tissues. The results indicate that the proposed extraction of rib bone efficiently isolated muscle tissues from linked liver tissues and improved the segmentation performance.

The existence of complex networks can be observed in various real-world contexts, such as social, biological, and/or neurological networks. A critical analytical challenge in such networks is community detection, which entails detecting groupings of nodes with dense internal connectivity. Numerous studies have been conducted on the subject of overlapping communities, wherein nodes may concurrently belong to multiple communities. In this paper, we propose an enhanced fuzzy clustering method for overlapping community detection based on neighborhood similarity. The core idea is to observe the community membership as a continuous feature; hence, nodes can belong to more than one community following different levels of affiliations. Our method consists of four stages: first, we find local structural features; then, we make a neighborhood similarity matrix based on common neighbors; next, we give initial fuzzy memberships using an Enhanced Fuzzy C-Means approach; and last, we improve memberships using a local optimization strategy. We evaluated our method on various real-world datasets of differing sizes and determined that it outperforms multiple state-of-the-art techniques, as indicated by overlapping modularity, F-score, and statistical significance assessments. The proposed method is a useful and scalable solution that is easier to understand and more accurate.

Phishing is one of the most persistent threats, but a lot of awareness programs still use generic, static training. This paper fills in the gap identified above by existing studies through the introduction of a phishing simulation platform that provides personalized, role-based simulation with real-time behavioral tracking. It is a multi-channel delivery (Email, Short Message Service (SMS), (WhatsApp) and can dynamically generate messages using placeholders to simulate realistic attack scenarios. User interactions are visualized on an integrated dashboard to let organizations judge the individual’s risk and provide immediate awareness feedback. Due to ethical restrictions, real user testing could not be performed, and the system was tested using simulated data found to work with a cloud-ready front end. The solution shows great potential for being adopted by enterprises due to its potential to adopt an approach towards cybersecurity training in a more adaptive and engaging way.

The escalating complexity of modern software systems has rendered the management of requirements increasingly arduous, often plagued by redundancy, inconsistency, and inefficiency. Traditional manual methods prove inadequate for addressing the intricacies of dynamic, large-scale datasets. In response, this research introduces SQUIRE (Semantic Quick Requirements Engineering), a cutting-edge automated framework leveraging advanced Natural Language Processing (NLP) techniques, specifically Sentence-BERT (SBERT) embeddings and hierarchical clustering, to semantically organize requirements into coherent functional clusters. SQUIRE is meticulously designed to enhance modularity, mitigate redundancy, and strengthen traceability within requirements engineering processes. Its efficacy is rigorously validated using real-world datasets from diverse domains, including attendance management, e-commerce systems, and school operations. Empirical evaluations reveal that SQUIRE outperforms conventional clustering methods, demonstrating superior intra-cluster cohesion and inter-cluster separation, while significantly reducing manual intervention. This research establishes SQUIRE as a scalable and domain-agnostic solution, effectively addressing the evolving complexities of contemporary software development. By streamlining requirements management and enabling software teams to focus on strategic initiatives, SQUIRE advances the state of NLP-driven methodologies in Requirements Engineering, offering a robust foundation for future innovations.

Interactive data exploration at scale remains constrained by 1) weak adaptability to shifting query workloads, 2) limited and post hoc error guarantees, 3) poor scalability under dynamic, high-dimensional data, 4) sparse user guidance during query formulation, and 5) non-trivial system overheads from learned or probabilistic components. We propose an end-to-end, privacy-aware framework that dynamically forms SQL queries for multi-dimensional data using randomized signals derived from personal web usage. The method integrates: 1) on-device user modeling that converts browsing interactions into preference embeddings under local differential privacy; 2) a constrained-randomization layer that enforces coverage and diversity to avoid filter bubbles while remaining responsive to user intent; 3) a contextual bandit policy (with optional deep reinforcement learning extension) that selects or completes query templates using signals from user profiles, session context, and data synopses; and 4) an error-aware AQP executor combining stratified/pilot sampling, synopsis reuse, and confidence-interval gating with automatic sample escalation. This design directly addresses the above limitations: the bandit adapts online to workload shifts; the AQP layer provides pre-execution feasibility checks and per-query error control; synopsis reuse and AB-tree–style random sampling maintain low latency under updates; and a guidance module (predictive autocompletion with information-gain scoring) reduces user effort while preserving exploration diversity. To evaluate effectiveness, we introduce a privacy-preserving training regimen (federated updates over DP-noised profiles) and a novel benchmark protocol measuring time-to-insight, error compliance under differential privacy, session diversity, and latency against strong baselines. The result is an ML-driven exploration loop that achieves error-bounded interactivity, robust personalization, and scalable performance on evolving, high-dimensional datasets, while providing evaluation metrics that capture both user experience and privacy-preserving guarantees.

Text mining methods often rely on a single data source or simple word frequency statistics, making it difficult to capture multi-source text semantic associations and local contextual dependencies, resulting in poor mining accuracy. Therefore, a method for text information data mining in natural language processing tasks is proposed. Using Python web crawlers to obtain multi-source text data, after preprocessing such as cleaning, segmentation, and removal of stop words, a Vector Space Model (VSM) is used for text representation, and a TF-IDF (Term Frequency Across Document Frequency) weight optimization mechanism is introduced to enhance feature semantic representation. On this basis, a semantic enhancement system is constructed based on the BERT classification model in the field of natural language processing. Through the self-attention mechanism of multi-layer Transformer encoders, semantics are aggregated to effectively capture local contextual dependencies, and context-sensitive word vectors are generated by the output layer. Finally, by fine-tuning the parameters of the BERT (Bidirectional Encoder Representation from Transformers) model and combining it with the Softmax function, precise mining of text information data categories was achieved. The experimental results show that in the embedding experiment of sports news headlines, this method can form a semantic aggregation structure with clear domain logic for word vectors; In the cross domain short text classification experiment, the overall accuracy of this method on the dataset reached 95.7%, which was 19.5% and 18.7% higher than the comparative methods, effectively solving the cross domain ambiguity problem in natural language processing.

With the ever-increasing rate of cyber threats, especially through malicious domain names, the need for their effective detection and prevention becomes very urgent. This study mainly investigates the classification of domain names into either benign or malicious classes based on DNS logs using machine learning. We evaluated five strong ML models: XGBoost, LightGBM, CatBoost, Stacking, and Voting Classifier, in an effort to obtain high accuracy, F1 score, AUC, recall, and precision. The challenge in that direction is to achieve a very good solution, without using deep learning techniques for low computational cost. Moreover, this project has an obligation to upgrade the cybersecurity landscape by embedding the best-performing model into the DNS firewall to enable protection against harmful domains in real time. Our dataset was collected and curated to include 90,000 domain names, including an equal number of safe and harmful, respectively, extracting 34 features from DNS logs and further enriched using publicly available data.

Electric Vehicles (EV) chargers rely on resource-constrained embedded hardware to execute critical charging operations. However, conventional security solutions may not adequately meet the needs of these devices. Increasingly, machine learning techniques are being leveraged to detect cyber attacks during electric vehicle charging. This study aims to evaluate various base machine learning methods and conduct binary and multi-class classification experiments to enhance security and operational efficiency in EV charging stations. The experiments utilize the CICEVSE2024 dataset, curated by the Canadian Institute for Cybersecurity at the University of New Brunswick, designed specifically for anomaly detection and establishing behavioral patterns in EV charging stations. The analysis highlights nuances in performance across different machine learning classifiers. For instance, Random Forest achieved 95.07% accuracy in binary classification by constructing robust decision trees. Ensemble methods such as CatBoost and LightGBM further improved binary classification to 95.37% and 95.41%, respectively through gradient boosting techniques. In multi-class attack classification, ensemble methods demonstrated superior performance, with the Stacking Ensemble achieving 91.1% accuracy by combining multiple models, and Voting Ensemble achieving 90.7%. Notably, among homogeneous base classifiers, Extra Trees and HistGradient Boosting were particularly effective, achieving 90.2% and 89.8% accuracy respectively in multi-class classification tasks. These findings underscore the efficacy of machine learning in enhancing cybersecurity measures for EV charging infrastructure.

Food preservation and safety require advanced detection methods to ensure transparency in supply chains. Terahertz (THz) spectroscopy has emerged as a powerful, non-invasive tool for material characterization. This study explores the integration of THz spectroscopy and machine learning for accurately quantifying maize starch adulteration in bioplastics derived from potato starch. Bioplastic samples with varying concentrations of maize starch were prepared, molded into three different thicknesses, and subjected to a two-stage drying process, resulting in 81 samples (27 treatments with three replicates each). The spectral profiles at THz (0.5 to 2 THz) were recorded and analyzed using three regression models: support vector regression, partial least squares regression, and multiple linear regression. The models were evaluated using the coefficient of determination (R2), Root Mean Square Error (RMSE), and the Residual Predictive Deviation (RPD). The results showed R2 values ranging from 0.7283 to 0.9495, RMSE between 0.0594 and 0.1393, and RPD values from 1.8753 to 4.4479, demonstrating strong predictive performance. These findings highlight the potential of THz spectroscopy and machine learning in the noninvasive detection of starch adulterants in bioplastics, paving the way for future research to enhance model robustness and applicability.