University of Bahrain Journals
The Deanship of Graduate Studies and Scientific Research was established by a decree from the Board of Trustees of the University of Bahrain to encourage scientific research and undertake contractual researches and studies for both public and private sectors in all fields. And with what the Kingdom of Bahrain is witnessing with respect to a serious desire to leverage the level of higher education and graduate studies in order to achieve its economic vision, it became mandatory for the University of Bahrain to focus on graduate studies to keep pace with those long-term aspirations and ambitious development visions. As such, and for the sake of the university to introduce a vision towards distinguished publication for the scientific production, the university has secured all the financial and moral possibilities to publish scientific production (whether authored or translated), in order to become a strong source for knowledge production and dissemination.
Each journal maintains its own specific policy, which can be found on the respective journal's webpage.
results
Discover
Communities
Select a community to browse its collections.
- This is a biannual peer reviewed journal. It publishes original research in the field of basic and applied sciences including Biology, Chemistry, Physics, Mathematics, Operation Research, Statistics, Astronomy and Geology. It also publishes specialized issues of relevant conferences and symposia held locally, regionally, or internationally. Production and Hosting by Elsevier B.V. on behalf of University of Bahrain http://www.elsevier.com
- This is a peer-reviewed International Journal that publishes technical papers describing recent research and development work in all aspects of digital system design and embedded systems.
- This is a semi-annual refereed journal. It publishes research on financial accounting, management accounting, auditing, taxation, social and environmental accounting and accounting education.
- This is a quarterly scientific refereed Journal specialized in educational and psychological studies. It publishes original academic research and studies in all educational and psychological fields that adhere to the commonly used scientific research methodology, and is written in Arabic or English
- This journal publishes research papers and specialized academic studies in the areas of Linguistics, Literature, Comparative Criticism, Philosophy and Human Thought, Sociology, Geography, Education, Arts, Folklore, Anthropology, and Archaeology.
Recent Submissions
Item type: Item , Access status: Open Access , An Interpretable Time-Respecting Attention-Based Seq2Seq Framework for Multi-Horizon Forecasting in Cyber-Physical Automation Systems(2025-12-5) Ahmad Dahe; Vladimir Valeryevich StuchilinPredicting the future is one of the key tasks of automation and cyber-physical systems. Deep learning methods suffer from biased performance estimates due to random data splits and low model interpretability. In this study we propose a time-respecting attention-based Seq2Seq model that generates interpretable multi-horizon forecasts and is tailored to non-stationary time series. The proposed framework includes two methodological advances: a stratified temporal partitioning algorithm to split data into training and testing sets, preventing information leakage. Also, an attention-improved encoder-decoder model that dynamically selects the most informative time intervals from historical data for making long-term predictions. The framework was tested on multivariate time series data from an automated process environment, and was compared against baseline long short-term memory (LSTM) and gated recurrent unit (GRU) models, with all models evaluated using same performance metrics at 6, 12 and 24 hours ahead. Proposed approach demonstrated stable performance, maintaining F1-scores above 0.95 across all horizons, and offered clear interpretability through attention and SHAP analyses that highlighted key temporal and feature dependencies. Overall, results prove that the time respecting Seq2Seq model enhances forecasting stability and explainability in cyber-physical automation systems and serve as a general framework for predictive analytics in intelligent industrial and IoT environments.Item type: Item , Access status: Open Access , Deepfake Technology: Threats, Detection Strategies, and Future Prevention Frameworks(University of Bahrain, 2025-12-02) Abdul-Munem Abdul-Hameed, Ayoob; Flaih Hassan, Nidaa; Department of Computer Science, University of Technology; IraqRecent advances in computer vision and deep learning have facilitated the emergence of Deepfake Technology, a technique capable of generating highly realistic yet artificially fabricated videos, images, and synthetic voices. Leveraging powerful architectures such as Generative Adversarial Networks (GANs), autoencoders, and recurrent neural networks, deepfake systems achieve a level of realism that renders manipulated content nearly indistinguishable from authentic media. The democratization of machine learning frameworks and computational resources has further reduced the technical barriers to entry, enabling both expert developers and amateur users to produce convincing deepfake content. While deepfakes offer potential benefits in fields such as digital entertainment, virtual reality, education, and assistive communication, their malicious use presents significant risks. Individuals may face identity theft or reputational damage; corporations and governments risk financial loss and misinformation campaigns; and broader social, political, and religious institutions may experience destabilization due to fabricated evidence or propaganda. This paper conducts a comprehensive review of the existing literature to examine the underlying mechanisms, applications, and implications of deepfake technology. It evaluates production techniques, detection methodologies, and the motivations of perpetrators, while also identifying current challenges in developing robust countermeasures. The findings reveal that deepfakes constitute a profound societal and security threat. However, the adoption of advanced detection systems, coupled with proactive regulatory frameworks and stringent legal measures, can substantially mitigate these risks and foster safer applications of this powerful technology.Item type: Item , Access status: Open Access , CSSAR: A Tri-Modal Semantic Alignment Metric for Boundary-Aware Code-Switching ASR Evaluation(University of Bahrain, 2025-12-02) Palivela, Hemant'; Narvekar, Meera; Department of Computer Engineering, D. J. Sanghvi College of Engineering (Autonomous), University of Mumbai; IndiaCSSAR improves Pearson correlation to 0.87, achieving a 30-point absolute gain over WER for code-switching speech recognition evaluation. Automatic speech recognition (ASR) for code-switching (CS)–the spontaneous alternation of two or more languages within a single utterance–remains difficult to evaluate, especially in low-resource settings. Widely used surface-form metrics such as word error rate (WER) and character error rate (CER) ignore meaning, treat all token positions equally, and over-penalize mixed scripts and rich morphology. We introduce the Code-Switching Semantic Alignment Rate (CSSAR), a linguistically informed metric that (i) grants graded credit to meaning-preserving substitutions through a tri-modal similarity function combining multilingual embeddings, phonetic edit distance and morphological overlap; (ii) magnifies errors at intra-word, clause-level and sentence-level switch points via prosody-aware boundary weights; and (iii) adapts to each language pair by optimizing four interpretable weights on a small development set. Evaluated on 2,474 Hindi-Marathi utterances, plus Mandarin-English and English-Spanish corpora, CSSAR achieves a Pearson correlation of 0.87 with human intelligibility judgements while cutting boundary-related errors by 31% and reducing penalties for valid semantic substitutions by 30%. Runtime remains practical at only 1.8 times WER thanks to cached embeddings and bitvector alignment. An accompanying visual-analytics toolkit further decomposes CSSAR into semantic, boundary, insertion and deletion components, offering actionable insights for CS-ASR system debugging.Item type: Item , Access status: Open Access , A Hybrid GAN-Augmented Deep CNN model for Brain Tumour Detection through the Eyes of Grad-Cam Visualization(University of Bahrain, 2025-12-02) Kaithwas, Abhijeet; Bankatrao Shinde, Snehal; Dhumale, Harshdeep; Shirsath, Pradyumn; Pikle, Nileshchandra; Borkar, Pradnya; Diwan, Tausif; Computer Science and Engineering, Indian Institute of Information Technology; Symbiosis Institute of Technology, Nagpur Campus, Symbiosis International (Deemed University); IndiaOne of the challenges in healthcare is the detection of brain tumors using MRI images. In this paper, the proposed approach involves fine-tuned models such as VGG16, ResNet-50, and DenseNet for classification, along with techniques to improve the visual interpretability of the results. Generative Adversarial Network is used to create synthetic images for dataset augmentation. It has been observed that the combination of GAN-based augmentation with optimized ResNet-50 results in superior performance, as it gives an accuracy of 99.12% on the dataset that has been enhanced. Additionally, Grad-CAM draws attention to locations that are significant to the tumor, providing visual reasons that are in line with radiological understanding. The suggested framework is positioned as a reliable and trustworthy instrument for computer-assisted brain tumor analysis as a result of the synergy between predicted accuracy and interpretability, which adds to better diagnostic confidence. Grad-CAM was used to make the results easier to understand by showing the parts of the images that the model used to make predictions. These visual explanations show where tumors are and help doctors to understand the main area, that makes them more confident in treatment. This work highlights the advantages of integrating optimal deep learning architectures with data augmentation and explainability tools to improve accurate and interpretable medical diagnosisItem type: Item , Access status: Open Access , MBDNet: A Hybrid Deep Learning Model for MRI-Based Neurological Disease Classification Using CNN, BiLSTM, and Vision Transformers(University of Bahrain, 2025-12-02) Balaso Khot, Pratima; Ravindra Patil, Meenakshi; Kumar G.N., Naveen; Research Scholar, Department of ECE, CMR Institute of Technology, Bengaluru, and Visvesvaraya Technological University; Research Supervisor, Professor, Department of ECE, CMR Institute of Technology, Bengaluru, and Visvesvaraya Technological University; Research Co Supervisor, Associate Professor, Department of ECE, CMR Institute of Technology, Bengaluru, and Visvesvaraya Technological University; IndiaMagnetic Resonance Imaging (MRI) underpins diagnosis of Alzheimer’s disease (AD), multiple sclerosis (MS), and brain tumors, yet manual interpretation is time-intensive and variable. This work presents MBDNet, a hybrid architecture that integrates convolutional feature extractors with Bidirectional LSTM and Vision Transformer blocks. Parallel MBConv, residual, and bottleneck pathways capture complementary local and global cues, while an atrous spatial attention module emphasizes diagnostically relevant regions. The spatial dependencies in disease pattern are extracted with atrous attention which provides better disease oriented features for improved classification performance. Evaluations on multi-disease MRI cohorts (MS, tumor, AD) show state-of-the-art performance: 98% accuracy (binary) and 96.3% (four-class), with consistently high ROC-AUC. Statistical testing confirms reliability (one-way ANOVA F being 669.45 and p less than 10e-18; Tukey HSD favors MBDNet over baselines). Explainability analyses using Grad-CAM indicate anatomically plausible attributions; faithfulness metrics report insertion AUC approximately 0.84, deletion AUC approximately 0.21, and a approximately 37% accuracy drop under weight randomization, evidencing dependence on learned parameters. The findings suggest a robust and generalizable framework for MRI-based disease classification that advances accuracy and interpretability. A balanced diagnostic accuracy is achieved by MBDNet with better interpretability on different comprehensive datasets with low computational overhead with 7.2 million learnable parameters. This makes model suitable on low configuration resources.