ABSTRACT. Road traffic accidents (RTAs) pose a significant global public health and economic challenge, with low- and middle-income countries like Bangladesh bearing a heavy burden due to their severity and frequency. Accurate prediction of accident severity is vital for enhancing road safety and informing proactive measures. This paper introduces a novel hybrid model that integrates transfer learning with TabNet, a deep learning architecture tailored for tabular data, to predict RTA severity. The model is developed using a real-world dataset from the Dhaka Metropolitan Police (DMP), enriched through advanced preprocessing techniques. Principal Component Analysis (PCA) reduces the dataset's 23 features to 18 principal components, effectively addressing data sparsity, high dimensionality, and class imbalance while filtering out noise to improve generalization. The model is further refined through hyperparameter tuning, optimizing its predictive performance. Experimental results demonstrate that this PCA-optimized and fine-tuned TabNet model achieves an accuracy of 86.8%, surpassing traditional methods, particularly in data-constrained environments. The approach leverages transfer learning to harness knowledge from related domains, enhancing reliability despite limited local data. Additionally, the model’s transparency is bolstered by SHapley Additive exPlanations (SHAP), which highlight critical factors like traffic control, weather, and lighting influencing severity predictions. This scalable solution offers a robust framework for real-time accident severity prediction, making it highly applicable for resource-limited settings like Bangladesh, and underscores the potential of modern machine learning to address pressing traffic safety issues in developing nations.

ABSTRACT. Social media is a valuable resource for information about disasters, which is essential for prompt action and well-informed decision-making in emergency situations. However, existing research predominantly focuses on English-language datasets, limiting applicability in low-resource languages such as Bengali. Addressing this gap, we introduce DisasterMDDB, a novel multimodal dataset comprising 6,330 Bengali social media posts annotated across six disaster categories, featuring both textual and visual data. Using this dataset, we evaluated various combinations of modality-specific pre-trained encoders to investigate their effectiveness in multimodal disaster classification. The best-performing model, XLM-R with ResNet50, achieved an accuracy of 95.90% and an F1-score of 0.96, significantly outperforming unimodal baselines and generic vision-language models. The public release of this dataset enables benchmarking and development of multimodal models tailored to regional disaster scenarios.

ABSTRACT. Mobile batteries face critical degradation challenges from temperature extremes (-20°C to 60°C), high-current fast charging (up to 4C rates), and cyclic aging, which significantly compromise device performance and lifespan. This study presents a rigorous comparative analysis of four commercial battery technologies: nickel-manganese-cobalt (NMC) lithium-ion, lithium-polymer (Li-Po), solid-state, and silicon-carbon (Si-C) under accelerated stress protocols simulating real-world conditions. Through controlled thermal cycling, fast-charging tests, and 500-cycle aging at 80% depth-of-discharge, we quantify degradation in capacity retention, internal resistance growth, and Coulombic efficiency. Results demonstrate that Si-C batteries provide 20% higher nominal capacity than NMC but exhibit 30% faster degradation at 45°C+ due to accelerated SEI growth. Solid-state batteries show exceptional thermal stability (0% thermal runaway incidence at 60°C) but suffer 25% faster degradation during 4C charging from interfacial delamination. Hybrid mitigation strategies, combining phase-change materials (PCMs), adaptive charging algorithms, and AI-driven management, synergistically reduce degradation by 55–60% across all technologies, extending Si-C cycle life from 350 to 800 cycles. These findings enable application-specific battery optimization: Si-C with PCM+AI for flagship devices, NMC with adaptive charging for fast-charging applications, solid-state with PCM for extreme environments, and Li-Po with PCM for cost-sensitive implementations, collectively enhancing sustainability and user experience.

ABSTRACT. Accurately predicting product prices in online marketplaces plays a vital role in maintaining a fair and competitive environment. For sellers it means setting prices that attract customers while maximizing profit. For buyers it helps ensure they’re paying a reasonable amount based on product value. As e-commerce continues to grow data-driven pricing has become increasingly important for informed decision-making on both sides. This study focuses on laptops, presenting a machine learning-based approach that estimates prices based on hardware specifications. 14 different regression models were evaluated including XGBoost, Random Forest and Support Vector Regression and developed a stacked ensemble combining Linear Regression, Ridge, RANSAC, Decision Tree and Histogram-Based Gradient Boosting. The ensemble outperformed individual models achieving a R² score of 0.8633. To make the model accessible and user-friendly this study deployed it as a web application using Streamlit, allowing users to input laptop specs and receive real-time price predictions. While this framework is tailored for laptops the approach is easily adaptable to other electronics offering a scalable, data-driven pricing solution for e-commerce platforms.

ABSTRACT. Solar energy is widely regarded as a highly promising source that is both renewable and sustainable. Accurate forecasting of global horizontal irradiance (GHI) is essential for improving the real-time operation and control of photovoltaic systems. Despite its importance, research on multi-step-ahead forecasting remains scarce. This study focuses on multi-step ahead GHI prediction using minute-level data collected from Bidyut Bhaban, the headquarters of the Bangladesh Power Development Board (BPDB), located in the Ramna area of Dhaka, Bangladesh. Forecast horizons of 5, 15, 30, and 60 minutes ahead are determined by utilizing several commonly used deep learning (DL) architectures. Model performance is assessed using mean absolute error (MAE), root mean squared error (RMSE), and coefficient of determination (R²). Experimental results show that BiGRU performs best for short-term forecasts (5 min: R² = 0.9306, 15 min: R² = 0.9047), while GRU excels for longer horizons (30 min: R² = 0.8862, 60 min: R² = 0.8634). Both models demonstrate strong potential for high-frequency solar irradiance forecasting in urban environments, offering accurate and computationally efficient solutions for grid operators in Bangladesh.

ABSTRACT. The rising incidence of weapons and fire-related security concerns in public and private places indicates a crucial challenge to safety, highlighting the immediate need for advanced intelligent surveillance systems capable of precise and automated object detection.. This study presents a comprehensive comparative analysis of various high-performance convolutional neural network (CNN) architectures, which have been used in multi-class visual threat object classification for detecting guns, knives, fire, and non-fire objects. These include, DenseNet121, VGG19, ResNet152V2, NASNetLarge, MobileNetV2, and InceptionV3. A custom dataset of 6,650 images was prepared, which was divided into training (70\%), validation (15\%) and test (15\%) sets to ensure equal representation of each class. Each model was initially initialized using pre-trained weights on ImageNet and fine-tuned using a custom classification head. The integrated training methods included data augmentation, normalization, and dropout regularization. Experimental results show that DenseNet121 outperformed all other architectures and achieved 95.08\% success in accuracy, precision 95.45\%, recall 95.08\%, and F1-score 95.27\%, respectively. Quantitative evaluation was further supported by confusion matrix and metric comparison plots. The sample predictions confirm that, DenseNet121 is highly reliable in detecting for all four classes in complex real life situations. This study highlights the importance of architectural design, such as dense connectivity and feature reuse to improve the performance and generalization ability of the model. This work establishes DenseNet121 as a highly effective backbone model for multi-class visual threat detection and lays the foundation for future extensions such as real-time implementation, video surveillance, and interpretable AI-based analytics.

ABSTRACT. In digital communication and authentication, text images such as scanned papers, official records, and legal certificates are essential. However, their integrity is being threatened by the growing availability of sophisticated image editing tools. In contrast to natural images, text images can have their meaning significantly altered by even small changes, like the removal or modification of a single word. We propose a Contrastive Language Image Pretraining (CLIP)-enhanced U Net architecture for identifying and locating text forgeries in document images in order to overcome this difficulty. The system creates a rich multimodal representation by combining high-level semantic representations taken from a pre-trained CLIP Vision Transformer with low-level spatial data from the original image. A symmetric encoder-decoder structure further refines these fused features for precise pixel-level segmentation of tampered regions. This combination of transformer-driven semantics and convolutional details enables the model to identify subtle and structurally irregular manipulations effectively. Experiments on the DocTamper-SCD dataset demonstrate strong performance, achieving an IoU of 0.64, F1-score of 0.74, precision of 0.92, and recall of 0.66, highlighting the potential of multimodal fusion for robust document image forensics.

ABSTRACT. The study of marine mammals is essential for understanding and protecting ocean ecosystems. Acoustic monitoring provides a non-invasive and efficient way to track their activity, but traditional methods often fail in noisy and variable underwater conditions. This work presents a deep learning framework for acoustic recognition of marine mammals using the VGGish model, a convolutional neural network pre-trained on largescale audio datasets. The model was fine-tuned to detect species specific acoustic signals after applying targeted preprocessing techniques to enhance feature quality. Rigorous training and validation on curated datasets produced a test accuracy of 95.61%, representing a clear improvement over conventional approaches. The framework is reliable, scalable, and suitable for integration into real-time detection systems to support conservation planning in regions with high acoustic activity. Future work will focus on expanding the dataset and enabling full real-time deployment to further strengthen its value for marine mammal research and biodiversity monitoring.

ABSTRACT. Globally, road accidents are major concerns, where driver fatigue, distraction, and unsafe behavior are the main causes of a significant proportion of incidents. Conventional driver monitoring systems mostly rely on simple sensors or rule-based methods that lack robustness under lighting variations or occlusion conditions. Despite CNN-based architectures showing promise in detecting drivers' behaviors, most models are computationally expensive in diverse driving situations. In this research, we propose an attention-enhanced deep learning (DL) architecture that can exhibit supremacy in detecting drivers' behavior under different circumstances. In total, 7,286 images of five distinct classes are utilized for training, validation, and testing purposes. Experimental results demonstrate that the proposed Eff-DANet can outperform other architectures with a 97.80\% F1-score. For better understanding, additional performance metrics are also considered for classifying the five classes of images. The t-SNE clustering shows that the model can distinguish the five classes precisely. This research is vital for minimizing road accidents, especially in underdeveloped countries, where driver monitoring systems are often limited and road safety measures are inadequate.

ABSTRACT. Virtual try-on technology has quickly developed into a game-changing solution in online fashion retail. It aims to reduce return rates and enhance customer satisfaction by providing realistic digital fitting experiences. However, many current methods have issues with visual misalignment, poor preservation of garment details, and limited practical integration into e-commerce systems. We proposed a high-fidelity virtual try-on system based on computer vision and deep learning, which combines OpenPose-based pose estimation, a U-Net-powered Segmentation Generator, a Thin Plate Spline (TPS)-based Geometric Matching Module (GMM), and an ALIAS generator for anatomically accurate image synthesis to overcome these challenges. Trained on the high-resolution Zalando-HQ VITON dataset, the proposed framework shows strong performance, achieving impressive segmentation accuracy with a dice score of 87.70 ± 1.35, IoU of 77.10 ± 1.45, and Pixel Accuracy of 90.00 ± 0.95. It also provides high perceptual quality, matching VITON-HD benchmarks, while preserving up to 80% of garment texture fidelity. A user study with 50 participants further confirmed the realism and fit of the garments, with an average satisfaction rating of 4.3 out of 5. In addition, Grad-CAM visualizations highlighted the model's focus on garment contours and texture-rich areas, while practical integration into an e-commerce platform showcased real-time responsiveness and usability. Overall, the system not only keeps the person's identity and the fine textures of garments but also proves to be responsive in real time, making it suitable for commercial e-commerce use.

ABSTRACT. For sensitive information, robust secrecy requirements are essential across commercial, procedural, and legal domains. Cryptography addresses this challenge but loses confidentiality upon decryption. This research proposes a dual-security approach combining encryption and steganography to enhance inaudibility, strength, and payload capacity. First, the Advanced Encryption Standard (AES) encrypts data using C# and a randomly generated 128-bit key, ensuring unreadability. Second, encrypted data is concealed within bitmap (BMP) graphics as video frames using steganography. The Least Significant Bit (LSB) method, integrated with triple Exclusive OR (XOR) and an advanced Prewitt pixel selection technique, ensures high imperceptibility and security during embedding. The proposed approach achieves superior data invisibility while maintaining frame quality. Statistical evaluations of quality metrics validate the method’s efficacy, confirming improved imperceptibility and robustness compared to conventional techniques. This novel integration of AES encryption and secure steganography offers a reliable solution for safeguarding digital content, meeting the stringent demands of modern data confidentiality.

ABSTRACT. Underwater Optical Wireless Communication (UWOC) has risen as a promising technology for high-speed, low-latency underwater data transmission. However, the performance of UWOC systems is significantly degraded by underwater turbulence, scattering, and absorption over long distances. This research employs a decode-and-forward (DF) relay to extend the communication range and improve link reliability. Both source-to-relay and relay-to-destination links are independently modeled using the mixture Exponential-Generalized Gamma (mEGG) fading distribution. This captures the combined effects of different underwater turbulences. Closed-form expressions for key performance metrics, including outage probability (OP) and bit error rate (BER), are derived for both heterodyne detection (HD) and intensity modulation with direct detection (IM/DD) schemes. The analytical results are validated through Monte Carlo simulations under varying environmental conditions. The findings show that the dual-hop structure significantly improves reliability and extends the feasible communication range in harsh underwater environments.

ABSTRACT. An ultra-wideband terahertz (THz) metamaterial absorber has been proposed, featuring a compact architecture and high functional performance. The structure is composed of a gold ground plane, a silicon dioxide (SiO2) substrate, and a patterned vanadium dioxide (VO2) top layer. By leveraging the phase transition of VO2, the design achieves tunability while maintaining polarization insensitivity and robustness against variations in incident angle. The absorber exhibits a broad operational bandwidth from 2.148 to 5.333 THz, corresponding to 3.185 THz, with two near-unity absorption peaks of 98.6% at 2.434 THz and of 97.4% at 4.808 THz. The use of durable materials ensures mechanical stability and manufacturability, addressing common limitations of existing absorbers. With its superior broadband response, tunability, and simple configuration, the proposed design offers a promising solution for next-generation THz sensing, imaging, and communication technologies

ABSTRACT. We present a high-efficiency photonic crystal fiber (PCF) biosensor that utilize the phenomenon called surface plasmon resonance (SPR). The sensor’s efficiency is numerically examined using finite element method (FEM). The sensor's straightforward circle lattice structure and chemically inert gold coating on the exterior guarantee robust plasmonic coupling and simple manufacture. The improve sensing capabilities of the design span the range of the refractive index from 1.29 to 1.38. Furthermore, it has been found that the sensor with a length between 37 µm to 462 µm provided the optimum balance between sensitivity and propagation loss, increasing the accuracy of detection. Our proposed sensor demonstrate resolution of 3.33 × 10⁻⁵ RIU, a sensitivity to amplitude 1313.59 RIU⁻¹ and wavelength sensibility of 3000 nm/RIU. The framework has an ultimate figure of merit (FOM) of 115.384 (1/RIU) and a minimum FWHM of 26 nm with a RI of 1.38. With its simple structure, wider sensing capability, and accurate resonance shifts, the proposed PCF-SPR sensor allows for the high-sensitivity detection of biological fluids (such as glucose and hemoglobin), solvents, and environmental contaminants.

ABSTRACT. In today's world, modern electronics require high computing power, portability, minimal power consumption, and also necessitate exceptional data stability and error endurance. Static Random-Access Memory (SRAM) is perfect for error-sensitive and high-stability systems because it retains data well for all noise, voltage fluctuations, and environmental stress. This paper provides a performance comparison of three CMOS based SRAM cells 10T, Low standby power 10T (LP10T), and Data Dependent Low Power 10T (DDLP10T) focusing on key performance parameters such as power consumption, static power, read and write delays, Static Noise Margins (SNM), total energy consumption, area, and Electrical Quality Metric (EQM). The LP10T cell demonstrates the lowest write power and read delay, at 9.361 nW and 792.756 pS respectively, while achieving the highest EQM of 19.12 × 10⁴² and superior stability with a Hold SNM (HSNM) of 233.33 mV and Write SNM (WSNM) of 436.73 mV. In comparison, the DDLP10T cell achieves the highest Read SNM (RSNM) of 187.5 mV and the lowest write delay of 13.288 pS, while exhibiting increased area and energy usage. Meanwhile, the 10T cell, with the smallest area of 0.692 µm² and lowest energy consumption of 0.111 pJ, shows lower noise margins and higher read/write delays. Overall, the LP10T design showed the most optimal performance among the SRAM cells. All simulations were conducted in EDA design tools using 22 nm CMOS technology.

ABSTRACT. Abstract—The comprehension of animal communication by humans remains a significant scientific challenge. Nevertheless, with appropriate training, limited interaction with certain bird species is achievable. This study proposes a novel technique for decoding avian communication, integrating artificial intelligence with advanced digital signal processing (DSP) techniques. The proposed approach is demonstrated through a case study focusing on the vocalizations of crows. This paper presents the study addressing first six steps. The study recorded and analyzed vocalizations from three contexts: solitary presence calls (Class A), group information sharing (Class B), and feeding/hierarchical interactions (Class C). Key acoustic features—frequency, bandwidth, and RMS amplitude were extracted using RavenPro software.

ABSTRACT. The escalating instances of weaponized threats within public and critical infrastructure areas demand real-time, accurate, and mass surveillance solutions. Traditional static camera-based systems are often beset with low field coverage, excessive detection latency, and poor adaptability to dynamic threat scenarios. To address such difficulties, this paper proposes a real-time monitoring system using UAVs with dual YOLOv11n deep learning models that conduct weapon and person detection in parallel and a threat-level classification module appended thereto. The proposed design employs an onboard Wi-Fi-enabled camera module on a quadrotor UAV to stream video to a Flask-based edge processing server for low-latency inference. The system is also augmented with spatial–temporal situational awareness via heatmap visualization, trend analytics of detections, and integrated dashboard–mobile application synchronization. A contactless charging station is integrated onto the UAV platform to enable automatic battery resupply without manual intervention; hence, mission endurance and downtime in persistent surveillance missions are maximized. Experimental testing on 5,000 annotated frames showed outstanding detection accuracy and robust threat classification (Safe: 98.7%, Low: 96.5%, High: 94.8%). Precision–Recall analysis yielded an AP of 0.938 for weapon detection, and ROC curves resulted in AUC greater than 0.95 for all classes, reflecting strong discriminative capability of the model. The system ran at real-time rates on edge hardware, achieving timely response capability. This UAV-based approach offers a scalable, mobile, and intelligent surveillance platform that can be used in perimeter defense, high-density event monitoring, and rapid response operations. Future work includes low-light enhancement, multi-drone coordination, and increased threat class coverage.

ABSTRACT. Proper development in urban cities with huge populations increases the number of cars, which naturally leads to severe traffic congestion, affecting the entire development of the city. Bangladesh faces different problems, and one of the most crucial ones is that of traffic offenses, which are causative factors for traffic jams and accidents, although there are proposed solutions, followed by an IoT-based smart traffic management system has been developed. Integrated tripwire laser sensors and cloud platform enable real-time violation detection and immediate enforcement responses. This system can smooth out traffic signals and hence optimize the overall flow by being integrated with existing traffic infrastructure. Real-time data has improved road safety and lessened congestion.

ABSTRACT. Early detection of skin cancer is critical for improving patient outcomes. This paper proposes an IoT-based portable system for automated skin lesion analysis, combining hardware and software solutions for accurate, real-time diagnostics. The system employs an OV7670 camera module with an Arduino UNO R3 to capture high-resolution images. Data is processed locally and transmitted securely via a multi-stage pipeline: Base64 encoding, HL7-compliant message formatting, RSA encryption, and MLLP-framed TCP/IP communication. For redundancy, images are concurrently uploaded to an FTP server, with MQTT enabling lightweight IoT server communication. Experimental results demonstrate robust performance, achieving 98.3% end-to-end accuracy with a mean transmission time of 6.8 seconds for 5 MB images. The system outperforms MQTT and CoAP in latency and reliability, maintaining sub-2-second response times across bandwidths (2G to broadband). Integrated AI classification (e.g., CNN models) enables automated lesion categorization (benign/malignant) within the ThingsBoard IoT platform. Storage optimization techniques ensure efficient handling of large datasets in the Master Database (MDB). Key innovations include (1) edge-level image acquisition with clinical-grade HL7 interoperability, (2) secure transmission via hybrid encryption and protocol framing, and (3) scalable architecture supporting 24–26 images processed per 30 seconds. The system’s portability, compliance with healthcare standards, and diagnostic accuracy make it viable for both clinical and remote settings. Future work will focus on enhancing AI models and expanding multi-modal data integration.

ABSTRACT. Dependability of electricity continues to present challenges in remote and displacement-affected areas of Bangladesh (Cox's Bazaar) in a humanitarian context. This paper describes a field-based model for decentralized solar energy deployment through a third-party financing mechanism, implemented in Camp 7 by Ulterior Engineering under CODEC in partnership with UNHCR. At Camp 7, a 3.5 kW hybrid off-grid solar photovoltaics (solar PV) system was deployed to power community services and operationalized within fourteen days. Key technical and institutional components of the pilot included a novel investment model, accelerated engineering implementation, shared-risk legal and formal agreements, and real-time performance verification. The paper covers the systems architectures, that is, Jinko 580 W solar modules with a 4 kW inverter and battery system with complete setup in the Tulips building and Fortune Solar 250 W solar modules in the Sunflower buildings, as well as initiatives for resilience, scalability, and community acceptance. In the main, low-resource settings do not have the required access to capital and operating budget; however, third-party ownership and humanitarian funding mitigate this gap. Performance metrics indicated that over 75% of daily consumption was met by solar, and reliance on fossil fuels now represents a tiny fraction of the operational costs. The model illustrates a replicable method for increasing clean energy in off-grid and displaced communities, with policy and engineering implications.

ABSTRACT. This work compares two one-shot image classification approaches: a Siamese network trained with binary cross-entropy loss and a ResNet-18 model pre-trained on ImageNet with a 1-nearest neighbor (1-NN) classifier. A combined model using concatenated embeddings from both networks with Principal Component Analysis (PCA) was also evaluated. Experiments on a modified MNIST dataset excluded one class during training to simulate unseen-class recognition. Pre-simulation analysis identified an optimal Siamese similarity threshold of -5.7. The ResNet-18 + 1-NN pipeline achieved 56.46% accuracy, while the Siamese network reached 99.74%, misclassifying only 3 of 8,991 negatives and 23 of 999 positives. The combined feature approach did not improve accuracy, indicating high redundancy between feature spaces. Results confirm the Siamese network's superiority for unseen class recognition and highlight the importance of metric learning in one-shot tasks. The findings also reveal that feature fusion offers no benefit when embeddings capture largely overlapping information.

ABSTRACT. Crop disease management is critical for global food security, but effective, large-scale implementation is hindered by the latency and labor costs of traditional detection methods. While deep learning offers promising solutions, a significant research gap exists in deploying models that are simultaneously accurate and computationally efficient for real-time, in-situ analysis on resource-constrained aerial platforms. This paper addresses this challenge by introducing an autonomous agri- cultural drone integrated with a single-board computer (SBC) for on-the-fly crop disease detection and geo-localization. Our primary contribution is the empirical validation of the YOLOv11s model’s superiority for this edge-computing application. Through a rigorous comparative analysis against established lightweight models such as YOLOv8n, YOLOv8s, and YOLOv11n. We evaluate performance based on mean Average Precision (mAP), frames per second (FPS), and model size. The results demonstrate that YOLOv11s achieves a state-of-the-art balance, offering substantial improvements in processing speed and a minimal memory footprint while maintaining competitive accuracy. This validated framework presents a scalable, cost-effective, and field-deployable solution that advances precision agriculture by enabling immediate, data-driven interventions to mitigate yield loss.

ABSTRACT. This paper proposes a high-efficiency solar water pumping drivetrain for rural irrigation in Bangladesh. The system integrates a Kalman Filter (KF)-based Maximum Power Point Tracking (MPPT), silicon carbide (SiC) switch-based power converters, and Direct Torque Control (DTC) for induction motor operation. Simulations in MATLAB/Simulink and hardware deployment on a TI C2000 microcontroller validate the approach. Compared to Perturb and Observe (P&O) and Incremental Conductance, the KF-MPPT reduces power losses to 5 W (vs. 38 W in P&O) and maintains 4.3 A output current. Additionally, SiC components reduce heat sink volume by 77%, enhancing system efficiency, stability, and scalability for off-grid applications.

ABSTRACT. Sentiment Analysis (SA) has become an essential task in natural language processing (NLP), enabling computational understanding of opinions expressed in text. Although Bengali ranks among the most spoken languages globally, it remains significantly underrepresented in sentiment analysis research. To address the scarcity of rich and fine-grained Bengali datasets, we present SentiFive a large-scale, manually annotated dataset consisting of 31,411 Bengali YouTube comments, labeled across five sentiment categories: নিশ্চিত নেতিবাচক (Strongly Negative), কিছুটা নেতিবাচক (Slightly Negative), নিরপেক্ষ (Neutral), নিশ্চিত ইতিবাচক (Strongly Positive), and কিছুটা ইতিবাচক (Slightly Positive). Over 400 native Bengali speakers contributed to the rigorous annotation process, ensuring reliability and diversity in sentiment representation. We benchmarked the dataset across binary, three-class, and five-class classification tasks using both traditional machine learning (ML) models and deep learning (DL) architectures. Experimental results demonstrate that performance declines as class granularity increases, with LSTM and BiLSTM models consistently outperforming ML baselines. Notably, LSTM achieved 74% accuracy for binary classification, BiLSTM achieved 59% for three-class, and LSTM achieved 39% for five-class classification. This work provides a foundational resource for advancing Bengali NLP and highlights the challenges and potential in fine-grained sentiment modeling for low-resource languages. We established performance baselines and intend to make the dataset “SentiFive” and the related code-base publicly available for future research.

ABSTRACT. Sentiment analysis for low-resource languages like Bangla presents a significant challenge, primarily due to the scarcity of specialized datasets and the linguistic nuances of colloquial text. To address this, we propose a novel three-layer stacking ensemble underpinned by a synergistic hybrid feature architecture. This core innovation fuses deep contextual embeddings from advanced neural models, including M-BERT and Bi-LSTM, with classical statistical term-importance scores from TF-IDF, creating a uniquely rich feature representation. Our model’s efficacy is validated on DramaSent-BD, a new, purpose-built corpus of 45,721 annotated user comments from Bengali drama forums, which we also introduce as a contribution to the research community. The proposed architecture achieves a state-of-the-art accuracy of 97.40%, establishing a new performance benchmark by significantly outperforming both standalone models and a non-hybrid stacking baseline. This work thus presents a dual contribution: a highly effective and interpretable model for Bangla sentiment analysis and a valuable public dataset to facilitate future research.

ABSTRACT. People of different ages frequently use technology for leisure activities, and gaming is a common pastime for many. However, playing video games may cause significant changes in behavior, both positive and negative. Research about those changes has been ongoing for a long time. Most of the research was conducted using sophisticated medical environments. We propose a multimodal CNN-based approach to observe emotional changes in players using facial and speech cues extracted from gameplay video frames. A vast amount of YouTube videos was collected from different online gaming streamers and then image and audio datasets comprising hundreds of those videos. Utilizing Facial Expression Recognition (FER) and Speech Emotion Recognition (SER) methodologies, our objective was to identify patterns of behavioral changes during gaming sessions and longitudinally. Multiple models were employed for both SER and FER. For FER, DenseNet121 was fine-tuned and achieved the best performance, and for SER, a custom CNN architecture, specifically optimized for speech features like MFCCs and spectrograms, was developed and outperformed other SER models. In our research, we established the effectiveness of our approach in discerning patterns associated with behavioral changes.

ABSTRACT. Social media platforms, despite their informal and iverse content, hold immense potential for the early detection of depression. This study introduces a hybrid deep learning architecture that integrates Multi-Head Attention mechanisms with RoBERTa and DistilBERT models to classify depression severity into three levels: mild, moderate, and severe. By combining RoBERTa’s strong contextual comprehension with DistilBERT’s lightweight and efficient sentiment analysis, the proposed system effectively captures both the semantic depth and emotional tone of user-generated text. Severity labeling is guided by sentiment scores obtained from the DT12the/distilbert-sentiment-analysis model. To enrich the training process, WordNet-based data augmentation is employed, enhancing linguistic diversity while preserving semantic integrity. Furthermore, the model incorporates sentence-level attention through parallel transformer pathways, promoting comprehensive feature integration. Optimized using early stopping and learning rate scheduling, the architecture achieves impressive accuracy rates of 96.50% on training data, 93.96% on validation, and 94.06% on the test set, highlighting its practical relevance for real-world mental health surveillance and intervention

ABSTRACT. Traditional cryptographic techniques that form the basis of blockchain security are becoming more vulnerable to quantum attacks as a result of the development of quantum computing. With an emphasis on Round5 algorithm, this paper explores the use of quantum-resistant cryptographic algorithms to strengthen blockchain security systems. Round5 operates as an all-purpose, secure, low-complexity cryptosystem with guaranteed security levels against both classical and quantum threats that will be found sustaining technological evolution in the future; therefore, without such an adjustment, all blockchain improvements will be at risk in the future. Firstly, our study involves an in depth understanding of how blockchain relies on cryptography for secure transactions and privacy. Second, it involves the potentially disruptive benefits of quantum computing and its ability to hack current cryptographic solutions. Finally, it involves the potential for ROUND5, a lattice-based cryptographic solution, to be integrated into a blockchain for better safety and operation. In this respect, we propose a quantum-resistant blockchain architecture to provide better safety without any drastic operational shortcomings. The quantum-resilient nature inspires confidence in this blockchain application for current and future technological use.

ABSTRACT. Warehouses face major safety risks such as fire, gas leakage and poor ventilation, which endanger workers and property. Existing systems are often limited to single-parameter monitoring. This work presents a Smart Warehouse Safety System that uses IoT (Internet of Things) technology for integrated, hazard detection and response. The system uses MQ-135 and MQ-2 gas sensors to detect gas levels and a DHT22 sensor to measure temperature and humidity, with all data processed by an ESP32 microcontroller. It carries out several safety actions including suppressing fire by activating a water pump, managing ventilation, operating a servo-controlled emergency exit, and triggering buzzer alerts. Through ThingSpeak, the system provides real-time monitoring, data logging, alert notifications and remote door operation through the TalkBack feature. Prototype testing showed response times under three seconds, reliable cloud communication and handled hazardous situations effectively.

ABSTRACT. The rapid growth of data center networks has made them prime targets for sophisticated and high-volume cyberattacks, particularly network intrusions. Traditional intru- sion detection systems often struggle to maintain accuracy and adaptability in dynamic traffic environments. To address these challenges, we propose a Hybrid Deep Ensemble Framework that intelligently detects and classifies network intrusions by leveraging the strengths of multiple deep learning models. Our framework integrates Convolutional Neural Networks (CNN), Bidirectional Long Short-Term Memory networks (BiLSTM), and Deep Neural Networks (DNN) to extract spatial, temporal, and abstract feature representations, respectively. These com- plementary features are then combined through a feature-level ensemble mechanism, followed by a fully connected classification layer. The model is trained and evaluated using a real-world dataset, ensuring robust performance in terms of accuracy, precision, recall, and false positive rate. Experimental results demonstrate that the proposed hybrid architecture significantly outperforms individual models and traditional machine learning approaches, offering a scalable and adaptive solution for secur- ing modern data center environments against diverse intrusion threats.

ABSTRACT. The Distributed Denial-of-Service (DDoS) attacks have been recognized as the most disruptive threat in cybersecurity, where the attacker produces a large amount of malicious traffic to the victim's network to make the system unavailable. DDoS attack mitigation is quite challenging due to IP spoofing, large traffic volume, and protocol exploitation in the network and transport layers. Our aim is to design a low-cost solution using open-source tools with deeper visualization, flexibility, and scalability, which are limited in most existing studies. In this paper, we have proposed a SIEM-based (Security Information and Event Management) architecture for the early detection of DDoS attacks to reduce the impact of service disruptions. We have incorporated important components such as Suricata (IDS), SIEM, and the Kibana Dashboard in this architecture. We have also analyzed the Kibana Dashboard with SIEM, which acts as a key feature for performance evaluation indicators. Using this method, we achieved a more insightful outcome for the detection of DDoS attacks on Layer 3 and Layer 4 of the OSI model.

ABSTRACT. Android stayed a lucrative target, so tools had to separate adware, banking Trojans, SMS malware, riskware, and benign apps. We used the CICMalDroid2020 dataset with more than 17000 apps collected from December 2017 to December 2018. We built a hybrid feature vector of length 564, combining 301 static features with 263 dynamic behaviors. We split the data into 70 percent training and 30 percent test. We tested seven classifiers, XGBoost, CatBoost, RandomForest, ExtraTrees, HistGradientBoosting, KNN, and SVM and compared three preprocessing schemes, original features, principal component analysis, and linear discriminant analysis. We measured accuracy, precision, recall, F1 score, and ROC AUC, with per class metrics and a surrogate decision tree for interpretation. Ensemble methods performed best. XGBoost reached accuracy 0.9747 and F1 score 0.9716 on original features, ExtraTrees and RandomForest trailed closely. PCA cut performance markedly, LDA kept more information yet still lagged the original features. KNN and SVM underperformed in all settings. Per class results showed near perfect SMS malware detection, while Adware and Benign had lower F1 scores. The surrogate tree highlighted package names and main activity counts as most influential, which attackers could spoof. Near perfect ROC curves for XGBoost suggested overfitting, external validation remained essential.

ABSTRACT. In this era of digitization, clinics are now connected to the Internet through the Internet of Medical Things (IoMT), enabling real-time monitoring of patient data. To ensure data security and patient safety, effective anomaly detection in IoMT networks is crucial and requires robust threat detection mechanisms. This study proposes a lightweight and efficient anomaly detection framework that utilizes a streamlined Autoencoder architecture, trained exclusively on benign network traffic from the publicly available CICIoMT2024 dataset. We introduce three novel metrics: Flag Intensity to capture aggressive connection patterns, Protocol Entropy to measure protocol usage random ness, and Traffic Density to identify temporal traffic anoma lies. Our model outperforms traditional methods (e.g., Isolation Forest, OCSVM) and achieves competitive results compared to state-of-the-art approaches. Moreover, we use SHAP (SHapley Additive exPlanations), to find out which features matter most and select 15 out of 45 original features along with the 3 additional features we introduced. Our proposed metrics (Flag Intensity, Protocol Entropy, Traffic Density) were among the top explanatory features. The proposed model achieved 97.6% accuracy and 0.990 AUC, demonstrating strong performance with minimal features. This work offers a robust, reliable, and efficient framework for zero-day attack detection in low-resource IoMT networks.

ABSTRACT. Corn leaf diseases threaten crop yield and food security in Bangladesh, where corn is a vital source of nutrition and income. Manual detection is slow and error-prone, while existing deep learning approaches often rely on computationally heavy pretrained models such as ResNet50 (192 MB), VGG19 (56 MB), and DenseNet121 (80 MB), which are impractical for lowresource agricultural environments. In this paper, we propose CornNetLite, a custom ultralight convolutional neural network for automatic classification of corn leaf diseases. The framework addresses six classes, comprising five major diseases- bacterial leaf streak, common rust, corn blight, gray leaf spot, and maize chlorotic mottle virus-alongside healthy leaves. Trained on a balanced dataset of 15,000 field and public images, CornNetLite achieves 98.93% accuracy and a macro F1-score of 0.98 on a 2,250-image test set, while requiring only 118 KB of storage. These results show that CornNetLite delivers state-of-the-art accuracy with ultra-low computational cost, enabling real-time deployment on mobile and IoT devices. By providing an efficient and scalable disea se monitoring solution, CornNetLite supports early intervention, yield protection, and precision agriculture in resource-constrained settings.

ABSTRACT. Oryza sativa, often known as rice, is a staple food for around half of the world’s population, especially in Asia and Africa. As a staple crop in South Asia, timely and precise classification of rice leaf diseases is essential to maintaining food security. This paper proposes an efficient and explainable ensemble framework for automated rice leaf disease classification using lightweight deep learning models. A comprehensive image dataset of seven rice leaf diseases was created by combining and preprocessing samples from various sources, then balancing the classes to correct dataset imbalance. We benchmarked com- pact architectures (MobileNetV3Large, EfficientNet-B0, RegNet-400MF, MobileViT-S) on a curated dataset, with RegNetY-400MF achieving 93.98% accuracy and a soft-voting ensemble reaching 94.18%. To enhance interpretability, we applied Grad-CAM visualizations to the CNN-based models, providing insights into the regions contributing to disease predictions. The proposed framework is highly efficient, generalizes well across data splits, and is readily adaptable for real-world agricultural deployment. This work offers a reproducible baseline for lightweight, scalable, and explainable crop disease diagnostics.

ABSTRACT. The purpose of the study is to analyze the effect of the AI implementation on the efficiency of the supply chain in the e-commerce industry in Bangladesh both the internal factors in the organization and the external environmental pressures. The study is based on the TOE framework and examines the impact of the AIL, DAC, TMS, ERT, MCI, and RCP on the performance of supply chains. Quantitative research methodology was adopted where convenience sampling was used to gather information about 130 valid respondents who had been involved in working in different positions in the e-commerce sector in Bangladesh. The hypotheses and structural relationships were tested by conducting the analysis of the data with the help of the PLS-SEM with SmartPLS 4. The results indicate that the strongest and most positive influence on efficiency of supply chain has DAC and TMS, but AIL, ERT, MCI, and RCP also play significant roles. These findings support the need to synchronize the level of technological implementation with the organizational preparedness and the external environmental requirement in a bid to streamline operations within the supply chain. In practice, the research proposes that companies should invest in AI not solely but also in the commitment of the leadership and training of the workforce to achieve their full potential in terms of efficiency. In social terms, a more efficient supply chain may be a way to increase customer satisfaction and minimize environmental harm by lowering waste and time loss. This study brings a new twist to the current body of knowledge as it contextualizes the effects of AI adoption on supply chains in an emerging economy, which has not been well researched. Though, there are such drawbacks as convenience sampling and cross-sectional data, which can influence the generalizability and causality. There is a possibility that future studies using longitudinal designs and sector specific dynamics can be conducted.

ABSTRACT. Agriculture has a meaningful contribution to both food and livelihood. Corn belongs to the major food crop in Bangladesh. However, its productivity is greatly hampered by leaf diseases—Common Rust, Gray Leaf Spot, and Northern Leaf Blight. The traditional manual diagnosis process is slow, laborious, and inefficient for farmers who do not have the necessary resources. To address these challenges, this paper proposes an automated, deep learning-based system for accurate and timely classification of corn leaf diseases, empowering farmers with scalable and affordable solutions. The PlantVillage dataset, comprising images of diseased and healthy corn leaves, is utilized for model development. Four state-of-the-art pretrained convolutional neural networks-InceptionV3, ResNet152V2, DenseNet201, and EfficientNetB4 are fine-tuned via transfer learning, achieving accuracies of 99.49%, 99.34%, 99.12%, and 98.76%, respectively. Additionally, a lightweight CNN is designed, with only three convolutional layers and no frozen layers. This model reached an accuracy of 95.76%. Although its accuracy is slightly lower, its compact design, faster performance, and lack of reliance on external weights make it perfect for offline use on mobile devices or edge devices in resource-limited rural areas. The results show that using transfer learning with deep CNNs achieves top accuracy, while the custom network balances performance, adaptability, low hardware demands, and interpretability. The proposed framework allows for early disease classification. This approach reduces yield loss and helps improve long-term food security in Bangladesh.

ABSTRACT. This study examined the key variables that impact the adoption of AI-based marketing tools in the e-commerce ecosystem in Bangladesh, with emphasis on the interrelationship between technological, organizational, and environmental aspects. A quantitative design of research was used and a structured survey questionnaire to 200 representatives of different e-commerce companies was conducted. Upon screening, 170 valid responses were tested in SmartPLS 4 to refer to the proposed model. The findings showed that PRA, TMS, OR, CP, and RE had a strong influence on the process of implementing AI marketing tools, with TMS the most significant one. Results also emphasized that employees with positive organizational culture and high leadership commitment made the introduction of AI successful. In practical sense, the research proposed that e-commerce firms ought to reinforce management, invest in employee capacity building, and enhance technological preparedness to enhance digital competitiveness. Although the study has a geographic limitation, it adds to the current body of knowledge concerning AI adoption in marketing because it includes empirical data and practical advice. Further observation is advisable in the future to explore the factors of competence of employees, cultural preparedness, and ethical aspects to gain more insights on sustainable AI-driven change in the emerging economies.