Our framework Phonesis is a machine-learned model of spoken language as an embodied mechanism over real human behaviour based on multimodal corpora (MOCHA-TIMIT, GRID, VoxCeleb2). Unlike approaches based on simulations, it combines a Speaker model (for interpreting the visual input and intention and converting it into realistic speech) and a Listener model (for the audio interpretation). Phonesis is trained in the end-to-end, which means that it is highly accurate in articulatory prediction, speech quality, and intent decoding. The results of ablation studies have indicated that joint optimization enhances performance. Zero-shot evaluation VoxCeleb2 has a high degree of generalization, indicating that it can be used in voice rehabilitation, brain-computer interfaces, and language understanding.
In the context of big data, balancing individual privacy with the need for data utility remains a critical challenge. This research presents the Bottom-Up k-Concealment (BU-KC) framework, a pioneering solution for Privacy-Preserving Big Data Publishing (PPBDP). At its core, the k-concealment model addresses the limitations of traditional k-anonymity by minimizing excessive generalizations, ensuring stronger privacy protection while preserving data utility. Combined with the Bottom-Up Generalization (BUG) strategy, BU-KC demonstrates superior performance compared to Top-Down Specialization (TDS) in terms of computational efficiency, scalability, and privacy preservation, while concurrently sustaining data utility. Furthermore, the integration of Apache Spark’s distributed computing paradigm enables us to effectively mitigate scalability constraints and processing bottlenecks commonly observed in the anonymization of large-scale datasets.
Cuckoo Hashing has been considered the state of the art since its origin. In Cuckoo Hashing, the search operation is executed sequentially in the tables, and thus, Cuckoo Hashing is Sequential Cuckoo Hashing. Based on the size of the table, Cuckoo Hashing is implemented in two ways: Symmetric Sequential Cuckoo Hashing and Asymmetric Sequential Cuckoo Hashing. Cuckoo Hashing suffers from higher insertion latency, inefficient memory usage, and high data migration. This work investigates the performance of both versions of Cuckoo Hashing under the light of two newly proposed performance indicators: Degree of Dexterity and Table Reference Count per key.
Creating non-invasive diagnosis procedures using speech is a very promising research area in biomedical engineering. Screening for an upper respiratory tract infection (URTI) or common cold using speech signals may be advantageous in terms of preventing its spread. In this study, we have proposed the various intrinsic mode multi-domain feature fusions for diagnosing the URTIs. First, each active speech frame is decomposed into several intrinsic mode functions (IMFs) using variational mode decomposition to capture the non-linear and non-stationary characteristics of pathological speech. Then spectral and entropy domain features are extracted from each IMF and used as features for classification. The URTIC and the newly recorded PRECRU URTIS database are employed to evaluate the efficacy of the proposed features. A transformer-based framework is employed to evaluate the discriminative capacity of the features, focusing on their capability to simulate long-range relationships with features. The proposed features outperform state-of-the-art methods, achieving a UAR of 70.82% and 72.24%, respectively, on the URTIC database's development and test partitions and 76.06% on the PREC-RU URTIS database. The results emphasize the efficacy of integrating intrinsic mode spectral and entropy features for reliable URTI identification
Social media and social networking have spread widely in everyday life, so it is important to use them in collaborative learning. Forming appropriate learning groups is, therefore, an important objective. This paper presents a novel approach based on a Genetic Algorithm (GA) for dynamic learners grouping in a Social Network Learning system (SNL). It offers some improved attributes applied for grouping learners and new genetic operators applied in the GA. The efficiency of the proposed approach was evaluated by comparing the groups formed using the proposed GA with randomly formed groups, resulting in the conclusion that the proposed GA is more effective and that the groups formed are more efficient.
Retinopathy of prematurity (ROP) is one of the notable causes of vision impairment among kids. The retinal zone, among the signs of ROP, is clinically considered a better predictor of severe forms than staging. This study explores five convolutional neural network (CNN) models used for computerized ROP zone classification based on fundus images of the retina. A total database of 1,365 images drawn from Al Amal Eye Center, located in Baghdad-Iraq, was trained into three classes mimicking three varied ROP zones. The images were used to finetune the models on top of pre-trained VGG16, VGG19, Xception, Inception-ResNetV2, and Inception-V3 models, whose network sizes and configurations were varied among the five models. Compared to minimal works done, the Inception-V3 model yielded the highest accuracy and reached 94.04% on zone detection. Computerized detection of the retinal lesions among pre-terms is fundamental to guiding the regimen plan that may incorporate laser therapy, intravitreal implantation, or close observation with intervention on demand. Merging computerized interpretation with the expertise of pediatric ophthalmologists may guide more consistent decisions on the management plan of the ROP. Future work will include external validation on independent multi-center datasets to assess generalizability.
Accurate segmentation of fine retinal vessels and the Foveal Avascular Zone (FAZ) in OCTA images remains challenging due to low contrast, noise, and class imbalance. To address this, a lightweight deep learning framework called VasFAZ-Net is proposed, which uses an EfficientNet-B0 encoder, a parallel spatial and channel squeeze-and-excitation (scSE) decoder, and a hybrid Dice-Focal loss function, trained on the OCTA-500 dataset with both 3 mm and 6 mm scans. The proposed model achieved Dice scores of up to 88.48% for vessel segmentation and 97.98% for FAZ segmentation with only 8.8 million parameters and an inference time of 18.2 ms per image. These results demonstrate that the proposed method provides accurate, computationally efficient, and clinically deployable multi-target OCTA segmentation for retinal disease analysis.
This paper introduces the 3x2A Strategy, a conceptual framework for societal adaptation to Generative AI (GenAI) and large language models (LLMs). It maps six strategic dimensions- Automation, Augmentation, Alliance, Alignment, Adaptation, and Accountability-critical for fostering human-AI synergy. We apply the framework to education, healthcare, and scientific research, illustrating its role in guiding responsible integration and governance. The 3x2A Strategy contributes to foundational discourse on AI alignment and sociotechnical transformation, offering a structured approach to aligning GenAI advancements with societal needs, institutional structures, and long-term resilience.
Understanding simply typed lambda calculus is essential for learning type systems and formal semantics, but its abstract concepts are often challenging. We developed a web-based tool that visualizes type checking through interactive proof trees, providing real-time feedback, highlighting typing contexts, and allowing step-by-step exploration of derivations. The language supports tuples, variants, lists, records, conditionals, recursion, and export of derivations in LATEX or image formats. User testing with university students showed improved comprehension and motivation. By making reasoning visual and interactive, the system connects theory and practical learning in type theory education.
Botnets continue to be one of the biggest cybersecurity risks since they provide a platform for a number of unlawful operations. The growing sophistication and stealth of contemporary botnet networks, which frequently elude conventional detection tools, make it difficult to identify botnets even with the availability of several detection approaches. A hybrid approach that analyzes both host-based activity and network traffic is suggested for detecting and neutralizing botnets. Tools such as Hping, Metasploit, Slowloris, and Hydra have been used to create a comprehensive dataset that blends both botnet and legitimate traffic, which helps to detect unusual activity. For efficient botnet traffic detection, the method uses a combination of the Gradient Boosting Algorithm (LightGBM) and Density-Based Spatial Clustering of Applications with Noise (DBSCAN) in an ensemble model after thorough feature extraction. A traffic neutralization technique stops the source IPs from trying to reach the destination after it has been detected. Explainable AI (XAI) techniques are incorporated to enhance interpretability and transparency, providing graphical visualisation and in-depth explanations of the identified botnet activity. A comparison study of some of the models, including Random Forest, Logistic Regression, Support Vector Machine, LightGBM, and DBSCAN individually, revealed that LightGBM combined with DBSCAN worked best with an accuracy of 95.6%. Furthermore, the ensemble model performed better than individual models by having 96% recall, 96% precision, and 96% F1 score. This comprehensive and uncomplicated method gives a strong and efficient botnet detection and neutralization solution, complementing current cybersecurity measures.
Brain tumors are the most grave and intricate diseases to be diagnosed and treated. In this study, a novel DCNN architecture is proposed. It has an end-to-end learning framework that automatically detects and refines relevant features directly from MRI images. The model will be trained and evaluated on a large-scale brain tumor MRI dataset that comprises SARTAJ, FIGSHARE, and BraTS datasets, which are the most widely used US brain tumor MRI datasets. In order to provide a fair and accurate performance evaluation, several transfer learning models and the concatenation of models will be utilized. The proposed DCNN outperforms pretrained models.
The degree of influence of data packet size on message latency in wide area computer networks is examined. Based on some queueing models and related calculations, it is shown that the influence in question is not univocal; there are many cases where segmentation is not appropriate even for relatively large message sizes. Accordingly, a methodology is proposed for determining the optimal size (in terms of minimum latency) of packets depending on factors such as: performance and load of routers and data transfer channels, bit error rate, number of ``source-destination" hops, etc.
The purpose of this paper is to analyse the challenges, identify the limits of traditional scoring systems, and propose solutions for selected ones to overcome them. Health and finance are generally considered the most critical areas of application of scoring systems, as scores can directly affect a person's life or quality of life. Other areas, for example, education, performance evaluation, human resources, justice, public safety, customer relationship management, marketing, and sports, are also important, but with less immediate effect. The domain of medical diagnostics was selected as the primary area for implementation, testing, and validation of the proposed solutions.
With the growing scale and complexity of log data, manual anomaly detection has become increasingly time-consuming and error-prone, necessitating the development of robust machine learning-based solutions. The HDFS (Hadoop Distributed File System) dataset, a large-scale real-world log collection, serves as a standard benchmark for evaluating both supervised and unsupervised anomaly detection methods. Available in two variants—a reduced and a complete version—this dataset facilitates comprehensive performance comparisons. Our empirical analysis reveals significant limitations in existing approaches: supervised methods exhibit poor performance on the reduced dataset, while most unsupervised techniques underperform across both versions. To address these shortcomings, we introduce several novel machine learning approaches for log-based anomaly detection. Additionally, we investigate the effects of alternative feature extraction techniques. We also examine the application of Synthetic Minority Over-sampling Technique (SMOTE) to mitigate class imbalance in supervised learning, as well as the incorporation of temporal features encoding inter-log time intervals. Our experimental results demonstrate that the proposed methods achieve statistically significant improvements in detection accuracy over existing approaches on both HDFS dataset variants, establishing new benchmarks for log-based anomaly detection.
This study presents a feature-level analysis of text complexity using large language models (LLMs) in a two-phase design. Phase syntactic complexity, coherence, named entities, and readability - achieving Spearman correlations of 0.55-0.60 across domains. Phase II employed indirect prompting to surface additional qualyielding a mean correlation of 0.42 and revealing that the six features account for 40% of complexity variance. Domain dependencies were limited to named entities and lexical diversity. We propose a hybrid model combining normalization, root-based synergies, and newly quantified metrics with domain-tuned formulae for improved prediction.
Processing texts from distant historical periods, especially those handwritten in languages with low computational resources, presents significant challenges. Even if modern methods make it possible to achieve, after laborious machine learning procedures, a fairly good rate of correct character recognition, the problem of the correctness of the resulting editable text remains a topical one. This paper presents an approach that contributes to the automation of the PostOCR proofreading process based on the presentation of digitized text using historical fonts, similar to those in the original document.
Gait, an individual's unique walking style, serves as an effective biometric tool for surveillance. Unlike fingerprints or iris scans, gait is observable from a distance without the subject's awareness, making it ideal for security applications. CNNs struggle with video variability, affecting gait recognition. This study introduces GaitDeep, a spatial-temporal refinement using a deep tion with a two-directional LSTM-based temporal module to prioritize key segments. Evaluated on the OU-ISIR, OU-MVLP, and CASIA-B datasets, GaitDeep achieves accuracies of 95.1%, methods and establishing a new benchmark for gait recognition.
An implementation of the Demoucron-Malgrange-Pertuiset ($DMP$) algorithm is proposed, based on specifying the notion of a segment (fragment, bridge) and developing an algorithm for calculating and updating the segments after each iteration using a depth-first search strategy ($DFS$). The algorithm also works for nonplanar undirected graphs by finally constructing a planar subgraph and displaying the list of segments that cannot be embedded, so as they generate edge intersections when drawing.
The proliferation of real-time, infinite data streams necessitates efficient online learning approaches. Hoeffding Trees (HT), which extend traditional decision trees using the Hoeffding bound, offer robust stream classification but face high computational costs. While the Green Accelerated Hoeffding Tree (GAHT) addresses energy efficiency concerns, its prediction accuracy can be improved by addressing its inherent limitations in combining Hoeffding bounds with information gain metrics for incrementally growing the tree. This study successfully develops enhanced GAHT variants through optimized Hoeffding bound stability and node splitting mechanisms. Our empirical evaluation demonstrates that the usage of these new variants improves predictive performance over the state-of-the-art GAHT, without compromising its energy efficiency.
The present research focuses on the security of Macedonian websites. It involves the analysis of HTTP Security header responses for 756 websites in the country, of which 246 are the most popular. This analysis is conducted across 13 different categories of websites, including government bodies and institutions, public institutions and enterprises, educational, commercial, news and media, entertainment, sports, etc. We intend to create a comprehensive security profile for the country's websites, which will help raise their overall security level. It is critical to understand and implement proper HTTP security headers to prevent or limit the dangers that can cause web-site attacks such as Denial of Service (DoS), Cross-Site Scripting (XSS), Cross-Site Request Forgery (CSRF), SQL Injection, click-jacking, etc. Our analysis was performed with the help of the Mozilla Observatory tool. We have discovered a significant lack of implementation and/or misconfiguration of HTTP security headers in all categories. Almost half of the websites (n=375; 49.60%) have an F grade, while more than a quarter of all websites (n=214; 28.31%) have a minimal security score of 0.
