Full Research Paper - NPS Australia Submission System

Towards Interpretable Legendary Pokémon Classification: A Comparative Evaluation of Machine Learning Models and Explainable AI

1. A comprehensive comparative evaluation of eight ML algorithms is conducted using a stratified five-fold Cross-Validation (CV) framework. The experimental findings demonstrate the superior effectiveness of gradient-boosting models, particularly LightGBM and XGBoost, in accurately classifying Legendary Pokémon and handling imbalanced gaming datasets.
2. A comprehensive ML framework is developed for Legendary Pokémon classification, integrating data preprocessing, feature scaling, missing value handling, and SMOTE-based class balancing to improve prediction reliability.
3. Integration of Explainable Artificial Intelligence (XAI) for Model Transparency:
SHAP and LIME are incorporated to provide both global and local explanations of model predictions, enabling transparent identification of the key attributes influencing Legendary Pokémon classification.

Supercapacitor-Based Energy Storage for Residential PV Applications: A Quantitative Comparative Study

Supercapacitors store energy electrostatically, have cycle lives of more than 10^6, and react to load transients in less than a second. All these properties have prompted several authors to suggest the use of supercapacitors as an alternative to electrochemical batteries in residential PV systems, which do not require maintenance. But the current literature is split, and methods that are proposed as fully battery-free for the supercapacitor have tended to be qualitative, without considering energy balance, conversion losses, or component lifetime, while methods that are proposed as simply using the supercapacitor as a transient buffer in a battery–supercapacitor hybrid have tended to be quantitative. The present work directly tackles this gap. A simulation of a 5 kW residential PV system, equipped
with a 40 kWh energy store, is created with one-second resolution, including capacitor self-discharge (leakage) and equivalent series resistance (ESR) and a voltage dependent buck–boost converter. The three storage architectures are analyzed: a capacitor-only store, a LiFePO4-only store, and a right-sized hybrid that integrates both. The capacitor-only architecture is demonstrated to be able to meet the load demand, but with a leakage of 3.4–3.9 kWh per day (17–20% of the load demand) and a round trip efficiency of only 68–86% versus 94% for LiFePO4; the bank also takes up 5–13 m^3. This is the limiting constraint in a levelisedcost analysis, where the maximum cycle life is constrained by calendar life, which gives the levelised cost of storage of 1.03– 4.12 USD/kWh – some 23 to 92 times that of LiFePO4. Either 68 equivalent full cycles per day or a cost of 109 USD/kWh of
the capacitors would be needed to break even with LiFePO4. In contrast, a small 0.3 kWh supercapacitor buffer (only 0.9% of the capacitor-only buffer size) embedded in the hybrid architecture is shown to reduce the current-ripple stress on the battery by 91%. These results suggest that the supercapacitor is not a viable bulk energy storage device, but could be applied in the power layer of a residential PV system where frequent shallow cycling could be economically justified.

Predictive Uncertainty for Mammography Classification Failure Detection: Limited Incremental Value of Saliency-Map Instability

This is the first controlled test of whether saliency-map instability adds failure-detection information beyond predictive uncertainty in mammography.

1. Across five CNN architectures and three settings (cropped-lesion CBIS-DDSM, multi-seed full-image CBIS-DDSM, and an external CMMD cohort), predictive entropy reliably detects classification errors (error-detection AUC 0.63–0.70) and supports effective selective prediction.
2. Using nested likelihood-ratio tests with paired bootstrapping and Benjamini–Hochberg correction, we show that saliency-map instability performs near chance on CBIS-DDSM (AUC approximately 0.50) and adds no significant information beyond predictive uncertainty in any full-image model, replicated across three seeds.
3. A controlled cropped-versus-full ablation shows the modest CMMD gains (AUC +0.015 to 0.054) are not reproduced on full-image CBIS-DDSM, pointing to dataset- or domain-specific factors rather than image format.

Development of an EfficientNetB0-Based Artificial Intelligence Model for Detection, Classification, and Separation of Polyethylene Terephthalate from Mixed Plastic Waste Streams

1. The paper proposes an EfficientNetB0-based six-polymer classification system for HDPE, LDPE, PET, PP, PS, and PVC, followed by practical PET/non-PET separation for recycling applications.
2. The paper improves dataset reliability through exact-duplicate removal, controlled oversampling, class weighting, augmentation, label smoothing, regularization, and multi-stage fine-tuning, reducing data leakage and class-bias problems.
3. The paper demonstrates strong PET-specific performance, achieving 96.84% accuracy, 94.85% precision, 92.47% recall, 93.65% F1-score, and 99.35% ROC-AUC on 1,109 independent test images.
4. The paper provides a practical pathway toward edge-based and industrial PET sorting, with future integration into cameras, conveyors, IoT monitoring, and robotic or pneumatic separation systems.

A Probabilistic Short-Term Load Forecasting with Conditional Diffusion and Split-Conformal Calibration Across Different Aggregation Levels

Probabilistic load forecasting requires reliable uncertainty estimates in addition to point predictions. This paper studies a conditional diffusion forecaster across national, regional, and household demand. The three datasets were preprocessed separately using the same 168-hour lookback, 24-hour horizon, and followed the same calibration pipeline. The diffusion model generated 50 trajectories for each forecast to form 90% prediction intervals. CNN-LSTM and N-BEATSx were used as point
forecasting baselines. N-BEATSx achieved the best point accuracy, while the diffusion model provided predictive distributions that the baselines could not produce. Forecast difficulty increased from 0.037 at the national level to 0.416 at the household level. However, the calibration failure did not follow this pattern.
The raw interval outputs of the diffusion model were under dispersed, with coverage from 0.631 to 0.808, so split conformal calibration was applied to restore the nominal 0.90 level. The split-conformal calibration improved coverage but remained below 0.90 for every dataset, whereas the block-random calibration restored coverage to 0.895–0.905. Panama showed the largest train-to-test mean temporal distribution shift (+0.3910) and the lowest chronological coverage of 0.843. These results
identify the directional temporal drift as a distinct failure mode of conformal calibration in chronologically partitioned load data. These findings are relevant to operational forecasting systems that rely on conformal intervals for reserve planning and risk assessment under non-stationary demand.

An Explainable Hybrid DeBERTa-BiLSTM-Attention Framework for Cyberbullying Detection In Social Media

Artificial Intelligence & Applications

Real Time Bangladeshi Vehicle Type Recognition Using YOLOv9 Variants

The paper develops and compares two YOLOv9 variants (YOLOv9-C and YOLOv9-E) for detecting 12 vehicle classes — including non-conventional local vehicles like rickshaws, legunas, easybikes, and wheelbarrows — using a custom annotated dataset collected from real Bangladeshi traffic signal points (402 images, 956 instances). This addresses the gap left by generic Western/East Asian detection datasets, which fail to generalize to Bangladesh’s heterogeneous traffic mix. YOLOv9-E achieved the best overall results (mAP@50 73.66%, precision 0.8508, recall 0.6433, F1 0.7327), while YOLOv9-C showed a marginally tighter mAP@50-95 (45.53%), highlighting an accuracy–localization trade-off between the two variants. The best-performing model was deployed via Streamlit to demonstrate real-time, practical feasibility for traffic monitoring.

AgroNeuro Guardian: An Integrated IoT-TinyML Framework for Autonomous Irrigation Control and Flood Prevention in Precision Agriculture

The AgroNeuro guardian is an IoT-based agricultural monitoring and controlling prototype that was designed and developed using ESP32, sensors, actuators. The experiment conducted above proves beyond reasonable doubt that the water-level sensor, LDR, PIR sensor, servo motor, relay, and 5V water pump were successfully operated. Indeed, the waterlevel sensor registered around 0% moisture level when it was kept in dry land and around 98% when it was kept in a wet condition. In addition, the LDR was successfully able to provide a particular digital output to operate the system. Most importantly, solar panels, TP4056 module, and a rechargeable battery were also successfully connected to the circuit. Thus, by and large, the prototype that was made is able to demonstrate the essential concept of agricultural monitoring and smart irrigation. However, there could be the following future implications, including the integration of Blynk remote monitoring system, cloud storage, intelligent irrigation, and machine learning for agricultural prediction.

Poultry Farm Micro-Climate Controller

The proposed IoT-based poultry farm micro-climate controller provides an integrated solution for real-time environmental monitoring, automated control, and remote supervision of poultry farm conditions. The system uses multiple sensors, including the DHT22, MQ-135, sound sensor, LDR, and BMP180, to monitor important environmental parameters. The ESP32 processes the sensor data and controls the connected actuators through relay modules according to predefined threshold conditions

A Hybrid ReliefF-Wrapper Feature Optimization Framework with Stacking Ensembles for Explainable Breast Cancer Diagnosis

The significant research contributions of this study are threefold: first, a two-stage Hybrid Feature Selection (HFS) strategy combining statistical filtering (ReliefF) with non-linear wrapper selection (Random Forest Gini importance) is implemented to prune the high-dimensional feature space down to the top optimal attributes, eliminating redundancy and reducing computational overhead; second, an optimized Stacked Generalization architecture fuses diverse $L_0$ base learners (Random Forest, XGBoost, and Support Vector Machine) with an $L_1$ Logistic Regression meta-learner, achieving superior predictive performance on the WDBC test set ($97.67\%$ Accuracy, $100.00\%$ Precision, $93.75\%$ Recall, $96.77\%$ F1-Score, and $99.83\%$ ROC-AUC); and third, a dual-level Explainable AI (XAI) framework integrating SHAP for global cohort feature attributions and LIME for local patient-level explanations resolves the “black-box” nature of ensemble models to provide transparent, clinically actionable diagnostic decision support.

SoC-Aware Adaptive Power Management for Multi-Source Energy Harvesting in Smartphone Charging

The growing energy demand of smartphones and other portable electronics has increased interest in supplementary energy-harvesting solutions that can operate when conventional charging infrastructure is limited. The integration of solar, radio-frequency, and kinetic-source based technologies can enhance energy availability; however static power-management strategies are not effective for time-varying sources. This paper proposes a state-of-charge (SoC)-aware adaptive power-management scheme which dynamically manages the three harvesting sources based on the available effective power instantaneously and the battery status. To determine source-allocation variables, a constrained Sequential Least Squares Programming (SLSQP) formulation is used to balance harvested-power utilization with consistency with respect to the prevailing source-power distribution. The framework is tested through simulations in the time domain, made with Python, using a 5000~mAh, 3.7~V battery under the same conditions, both for static and adaptive strategies. The adaptive method can shorten charging time from 13.02~h to 10.98~h, increase the average charging power from 1.13~W to 1.35~W, and reduce the unused energy from 21.15~Wh to 9.53~Wh and increase the energy utilization from 41.09\% to 60.72\%. Further simulations under cloudy, low-light and low-energy conditions indicate that the relative gain of adaptive allocation becomes more pronounced as harvested-power availability becomes more restricted. The results show that battery-aware dynamic source coordination can enhance the effective utilization of harvested energy from multiple and heterogeneous sources in time-varying operating conditions.

Coordinated Vehicle-to-Grid Control for Smart Microgrids: A Real-Time SOC-Aware Strategy for Frequency Regulation and Grid Stability

There is an absence of a thorough, well-coordinated control plan for microgrid integration of Vehicle-to-Grid (V2G) technology. Electric vehicles (EVs) with V2G capabilities have the potential to stabilize the grid, but their practical application is constrained by the inadequacies of present models in handling their real-time interactions. This creates a significant knowledge gap regarding how Vehicle-to-Grid (V2G) technology can enhance the overall performance, stability, and dependability of microgrids. Designing and analyzing a system that can efficiently control EVs is imperative to handle issues like load balancing, frequency regulation, and state-of-charge management. In order to address the challenge of integrating Vehicle-to-Grid (V2G) technologies, this study proposes a Simulink-based microgrid model. When the AC grid frequency drops to 0.998 p.u. at 10s and 30s, the battery lowers its charging current from roughly 2000 A to 1500 A and its DC-link voltage from over 540 V to 535 V to reduce the grid’s demand, demonstrating the G2V system’s capacity to operate as a responsive load. However, the V2G system functions as a power source, increasing its active power output from 0.5 MW to 1 MW to offset the same frequency dips. In order to balance loads and stabilize the grid, an innovative control approach regulates EV charging according to the EVs’ current state of charge. According to the study’s simulations, this V2G-enabled system works noticeably better than conventional setups, demonstrating the technology’s viability and advantages for smart microgrids.

An Ensemble Vision Transformer Approach for Automated Detection of Hepatocellular Carcinoma

Hepatocellular carcinoma (HCC) is the most common type of primary liver cancer in adults and the third leading cause of cancer-related mortality worldwide. Early-stage detection and diagnosis are crucial for reducing patient mortality. The conventional clinical standard relies on histopathological image examination, which requires substantial manual effort, extended processing time, and is prone to observer-dependent variation. Advanced machine learning (ML) and deep learning (DL) approaches have been widely applied to support early cancer detection and diagnosis. However, the scarcity of large annotated medical image datasets and reliance on single models may degrade performance, limiting the development of robust automated systems.To address these challenges, this study proposes a robust classification framework to distinguish HCC from healthy liver histopathological images. The proposed pipeline integrates data preprocessing and augmentation, deep feature extraction using ViT-B/16, LASSO-based feature selection, and a calibrated weighted soft voting ensemble classifier combining Logistic Regression (LR), Support Vector Machine (SVM), CatBoost, and XGBoost. The ViT-B/16 model effectively captures both local and global contextual dependencies through multi-head self-attention mechanisms, while LASSO performs simultaneous feature selection and regularization through cross-validation penalization. The final ensemble aggregates the decision boundaries of multiple classifiers and achieves an accuracy of 96.62%, AUC of 99.56%, precision of 96.62%, sensitivity of 96.57%, specificity of 96.67%, F1-score of 96.62%, and MCC of 0.9324, outperforming existing state-of-the-art methods. These findings demonstrate the effectiveness of integrating transformer-based feature extraction, interpretable feature selection, and ensemble classification for improving clinical decision support in liver cancer diagnosis.

An Explainable Hybrid Deep Learning Framework for Retinal Disease Classification Using DenseNet201 and SHAP-Optimized xLSTM

Retinal diseases such as diabetic retinopathy (DR), age-related macular degeneration (AMD), and cataracts are among the leading causes of blindness and visual impairment worldwide. These conditions often progress silently without noticeable symptoms during the early stages, making timely and accurate diagnosis crucial to prevent irreversible vision loss. Retinal fundus imaging provides a non-invasive approach for capturing structural abnormalities in the retina and plays a vital role in automated disease screening. In this study, we propose a multi-class retinal disease classification framework using fundus images. The proposed methodology integrates deep feature extraction using DenseNet201, feature selection using SHAP to identify the most informative features, and an xLSTM classifier for final prediction. Experiments were conducted on the AMDNet23 dataset, which contains 2,000 curated fundus images across four balanced classes: Normal, Diabetes, Cataract, and Age-related Macular Degeneration. The proposed configuration achieved an accuracy of 92.00%, an AUC of 0.9893, and an MCC of 0.8958, outperforming other evaluated model combinations. These results demonstrate the effectiveness of combining deep feature learning with interpretable feature selection for retinal disease detection and highlight the potential of the proposed framework as a reliable clinical decision-support tool for automated retinal screening.

Critical Success Factors for Cloud-Based ERP Adoption in Manufacturing Industry: A Systematic Literature Review

This systematic literature review (SLR) aims to identify, classify, and analyze the critical success factors (CSFs) that influence cloud ERP adoption in manufacturing industries. Adhering to the Preferred Reporting Items for Systematic Reviews and MetaAnalyses (PRISMA) 2020 protocol, a comprehensive search was conducted across five major academic databases: Scopus, Web of Science, IEEE Xplore, ScienceDirect, and Emerald Insight, resulting in 1,247 initial records. Following the application of
stringent inclusion and exclusion criteria, 24 studies were selected for final analysis. Utilizing the Technology-OrganizationEnvironment (TOE) framework as the analytical lens, this study identified 26 distinct CSFs. The findings indicate that
organizational factors dominate, accounting for 47.7% of total occurrences. Specifically, top management support (87.5%), change management (70.8%), and vendor support (66.7%) emerged as the most critical factors. Data security (54.2%) represents the most significant technology-related concern. Geographically, Asian countries contributed 45.8% of the research output, with a notable acceleration in publications observed post-2020. These findings provide valuable insights
for manufacturing managers, ERP vendors, and policymakers seeking to enhance cloud ERP implementation success.

Defect-Aware Hybrid Attention and Anchor Optimization for Automated PCB Defect Detection in Industrial Inspection

This work proposes a defect-aware Faster R-CNN framework that combines residual CBAM feature refinement with dataset-informed anchor configuration for small PCB defect detection. The model achieved 95.57% mAP@0.5 and 95.59% F1-score while operating at 23 FPS, demonstrating a practical balance between detection accuracy and inference efficiency.

Explainable Hybrid Deep Learning for Tuberculosis Detection in Resource-Constrained Settings

Tuberculosis (TB) remains a significant global public health challenge, particularly in resource-constrained settings where limited access to specialized diagnostic services may delay disease detection. This study developed and comparatively evaluated an automated and explainable framework for TB detection from chest radiographs using conventional machine learning, deep learning, and hybrid learning approaches. A publicly available Kaggle dataset comprising 4,200 chest X-ray images, including 3,500 normal and 700 TB-positive cases, was used for model development and evaluation. The images were resized, normalized, augmented, and balanced to improve training consistency and address class imbalance. Conventional classifiers were trained using Histogram of Oriented Gradients (HOG) features, while seven deep learning architectures and multiple CNN-based hybrid models were evaluated using 70:30 and 80:20 train–test splits under different augmentation and optimization settings. Among the conventional classifiers, SVM achieved a maximum accuracy of 99.50\% with data augmentation. VGG16 achieved the highest overall deep-learning accuracy of 99.88% under the 80:20 non-augmented configuration, whereas DenseNet201 achieved 99.86% with data augmentation. Among the hybrid approaches, VGG16 combined with Logistic Regression achieved the highest accuracy of 99.76\%. Gradient-weighted Class Activation Mapping (Grad-CAM) was further employed to enhance model interpretability by highlighting image regions that contributed to classification decisions. The findings demonstrate that deep learning and hybrid approaches provide highly competitive performance for automated TB screening, while Grad-CAM enhances the interpretability and transparency of model predictions. However, external multicenter validation is required before clinical deployment.

Deep Feature Extraction with Principal Component Analysis for Congenital Heart Disease Classification from Chest Radiographs

This study proposes a reproducible hybrid framework for four-class congenital heart disease classification from chest X-rays by combining ImageNet-pretrained ResNet50 and EfficientNetB3 feature extraction, PCA-based dimensionality reduction, and six conventional machine-learning classifiers. The study provides a controlled comparison of deep feature representations and heterogeneous classifiers under a consistent multiclass protocol. EfficientNetB3–PCA–SVM achieved the best performance with 93.82% accuracy, 0.938 macro-F1, 0.979 specificity, and 0.9947 macro ROC-AUC using only 91 principal components, demonstrating that compact PCA-reduced deep features can provide accurate CHD classification while reducing feature dimensionality and computational complexity.

AI-Driven Air Quality Analytics and Decision Support Framework for Dhaka City

This work is among the first to combine AI-based ward-level air quality prediction with an actionable, tiered decision-support and control layer specifically for Dhaka — closing the gap between just 3 physical monitoring stations covering 111 wards, and translating raw predictions into concrete, low-cost, institutionally realistic intervention priorities rather than simply reporting pollution numbers.

More specifically, the contribution operates on three levels:

Technical contribution: Combines sparse ground-truth data (3 CAMS stations) with satellite-derived aerosol data (Sentinel-5P, MODIS) and an LSTM-based spatiotemporal model to estimate hyperlocal AQI across all 111 wards of Dhaka — extending air quality visibility far beyond what existing station-level or city-level reporting can offer.
Decision-support innovation: Moves beyond prediction alone by introducing a rule-based priority-tier system (Critical/High/Moderate/Safe) that maps each ward to specific, evidence-grounded control actions — informed by real Bangladesh-specific field evidence (Brooks et al., 2024 RCT on brick kilns), making the recommendations institutionally realistic rather than idealized.
Scalability and replicability: Relies entirely on publicly available data with no new hardware deployment required, making the framework low-cost and directly transferable to other pollution-burdened Bangladeshi cities (Narayanganj, Gazipur, Chattogram) and comparable South Asian urban contexts.

A Secure Architecture for Immutable and Verifiable OTA Firmware Updates for IoT Devices using Blockchain

The rapid growth of Internet of Things (IoT) devices has intensified the need for secure and reliable over-the-air (OTA) firmware updates. Ensuring the confidentiality, integrity, and authenticity of firmware during updates is critical to prevent device compromise and maintain trust. However, existing systems still face challenges in providing fully secure OTA firmware updates. This paper proposes a blockchain enabled architecture that ensures immutable, verifiable, and confidential OTA firmware updates for IoT devices. The framework employs smart contracts to manage firmware metadata such as version, timestamp, and cumulative hash stored immutably on the blockchain, while the actual firmware binary is securely maintained in off chain cloud storage. To support lightweight verification on resource constrained devices, a chunk based cumulative hashing mechanism is introduced, where firmware is segmented into 100 byte blocks and aggregated using an XOR based hash computation. The system was implemented and tested on the Ethereum Sepolia testnet and validated on an ESP32 IoT board. Experimental results confirm secure update verification with minimal computational overhead. By integrating decentralized trust, tamper resistance, and efficient integrity validation, the proposed framework offers a scalable and transparent solution for secure OTA firmware management in IoT ecosystems.

An Empirical Evaluation of Quantum Kernel Stability in Network Intrusion Detection

This research shows that quantum kernel classifiers can become unstable even without hardware noise. It identifies seed-dependent classifier failures and links this instability to limitations in the quantum kernel representation. The findings also highlight the importance of reporting seed variability, worst-case performance, and class-specific metrics rather than relying only on average accuracy.

Performance Evaluation of a RIS-Enabled Hybrid THz/RF System Using Hard and Soft Switching for Next-Generation IoT Networks

This study presents a novel RIS-enabled hybrid THz/RF communication framework with hard and soft switching strategies to support the diverse connectivity requirements of future IoT networks. Closed-form analytical expressions for outage probability and average bit error rate are derived, while the results demonstrate the superior performance of soft switching under various fading conditions, highlighting its capability to provide reliable, adaptive, and high data rate connectivity for next-generation IoT applications.

OsteoNet: Gated Multi-Scale Channel Calibration for Binary Knee Osteoarthritis Screening

OsteoNet integrates a ConvNeXt-Large backbone with a Gated Multi-Scale Channel Calibration Module (MSCCM) for binary knee osteoarthritis screening using a KL0–1 versus KL3–4 split, achieving 97.56% accuracy and an AUC-ROC of 0.9946. The framework is further deployed as a containerised web application with Grad-CAM visualisation, demonstrating end-to-end inference and model interpretability beyond offline evaluation.

Analytical Degradation-Reserve Sizing of PV–Battery Systems Considering Lifecycle Aging and Service-Equivalent Battery Replacement

This research presents a closed-form degradation-aware sizing framework for PV–battery systems that clearly separates baseline capacity from the additional reserve required for aging. It introduces a service-equivalent comparison between battery reserve and replacement strategies under the same usable-energy requirement. The framework provides analytical expressions for degradation reserve, replacement timing, lifecycle cost, and economic break-even time. Its robustness is further evaluated using an alternative aging model and sensitivity analysis.

Empowering Smallholders: An SDG-Aligned Adaptive MCDM Algorithm for Agricultural Service Selection in Bangladesh

The significant contribution of this research is the development of a farmer-centric, adaptive decision-support algorithm that directly empowers Bangladeshi smallholders to select agricultural service providers, bypassing traditional informal broker networks. By integrating Haversine spatial filtering with a user-configurable TOPSIS multi-criteria evaluation, the system allows farmers to transparently rank providers in real-time based on customized weights for price, distance, service quality, and availability. Ultimately, this tool fills a critical gap in digital agriculture by providing a scalable, field-ready architecture that democratizes market access while simultaneously advancing multiple Sustainable Development Goals.

Machine Learning Based Heart Disease Prediction via Multi-Dataset Harmonization and Structural Validation

This study addresses the critical limitations of dataset fragmentation and poor generalizability in cardiovascular machine learning by harmonizing three disparate public heart disease datasets into a unified 2,271-record benchmark with a standardized 14-feature schema. Rather than relying solely on predictive metrics, this work pioneers a post-integration structural validation approach using Pearson correlation analysis to statistically prove that the merged dataset maintains clinical coherence and reduces source-dependent feature bias. Furthermore, rigorous cross-dataset benchmarking demonstrates that models trained on this harmonized dataset—led by a Random Forest classifier achieving 85.27% accuracy and a 0.910 ROC-AUC—empirically outperform identical models trained on isolated baselines. Finally, the framework ensures medical transparency and clinical plausibility by augmenting traditional ensemble methods with a complementary Graph Neural Network (GNN) extension and SHAP-based explainability to map decisions directly to key physiological predictors.

“Can AI replace human connection?” A study exploring human connection and emotional companionship.

This study contributes by distinguishing AI companionship as a source of emotional support from actual preference for replacing human relationships.

CUDA-Based Genetic Algorithm for Set Union Knapsack Optimization with Hybrid Initialization

This paper proposes a CUDA-based hybrid genetic algorithm framework that combines heuristic and random initialization with mutation, crossover, and parallel local search to efficiently solve the SUKP. The framework mainly uses heuristic initialization to generate high-quality solutions while using a small proportion of randomly generated solutions to maintain population diversity. Specifically, 20% of the elite population is used for heuristic initialization, with 99.5% of the GPU threads starting from heuristically generated solutions and the remaining 0.5% starting from randomly generated solutions. A mutation operator is then applied to improve population diversity and explore the search space. Each GPU thread independently performs parallel local search, while the crossover operator enables information exchange among individuals by using the best solution identified within each GPU block.

Renewable Energy Adoption as a Catalyst for Business Value Creation: Evidence from Manufacturing Firms in Bangladesh

The study added a theoretical contribution by showing that REA, EMC, TR, and GIC can serve as strategically relevant resources and capabilities to create BVC, and that the unimportant EO–BVC link indicates that environmental orientation alone can not be enough to be a resource or capability for BVC. In practical terms, the results advocate for manufacturing managers to pay attention to energy-management capabilities, technological readiness and green innovation, instead of viewing renewable-energy adoption as a standalone investment. Empirically, the study presents evidence from the manufacturing sector, which is relatively underrepresented in the renewable-energy and business-value research fields in the context of Bangladesh. The most powerful impact of GIC is to recognize the need for converting environmental efforts into innovation oriented products, processes and practices. The novelty of the study is that the REA, EMC, TR, EO and GIC are combined in a single RBV-based framework to explain the BVC, thus linking renewable-energy adoption with organizational capabilities and value creation.

BioOrbitX: An Explainable Cross-Tissue Framework for Prioritizing Spaceflight-Responsive Genes

Spaceflight perturbs physiology across organs, yet candidate-gene prioritization is commonly tissue-specific or based solely on differential-expression (DE) statistics. This study examines liver (OSD-168; 31,428 genes; n=5 ground-control and n=5 spaceflight animals in the selected RR-1 contrast) and soleus muscle (OSD-104; 22,437 genes; n=6 and n=6) RNA-seq contrasts. Genes with Benjamini–Hochberg-adjusted p1 were designated DE responders. Random forest, XGBoost, L1-logistic regression, and a soft-voting ensemble were evaluated within tissue and transferred between tissues using six expression/topology descriptors. Single-split AUROC/F1 estimates reached 0.934/0.443 in liver and 0.978/0.534 in muscle, whereas transfer was weaker (best AUROC 0.799, liver-to-muscle). SHAP ranking identified tissue-specific candidates, including Smad3, Myorg, Arrdc3, Bdh1, Fzd9, Mettl21c and Nqo1 in muscle; the top-50 sets did not overlap. An implementation audit showed that the original column filter retained source-table summary/statistic fields alongside per-animal columns; therefore, the reported model scores are exploratory and not a count-only, leakage-free validation. The contribution is a fully specified candidate-prioritization workflow and an explicit reproducibility boundary for the required count-only, repeated-resampling rerun.

BC-LMF: A Leakage-Controlled Multimodal Fusion Framework for Breast-Cancer Hormone-Receptor

BC-LMF paper revised: abstract reformatted to Background/Methods/Results/Conclusion at 250 words; added a state-of-the-art comparison table to Related Work; added the proposed model’s mathematical justification (6 equations) to Methodology; stripped other-model mentions from Results; cut Future Work to one item; removed commas before “and” and tightened prose. Verify page count and table labels on Overleaf.

Broken Tokenisers and Batch-Ordered Corpora: An Audit of Bangla Religious Hate Speech Detection

A silent tokenisation failure in the standard Bangla pre
processing pipeline is quantified: 83.2% of vocabulary
types are discarded at a cost of 0.0873 macro-F1, and a
one-line correction recovers them.
• Runs and rank-correlation tests establish that the corpus
is batch-ordered rather than random, and the resulting
evaluation inflation under naive splitting is measured.
• The construct validity of the religious hate label is ques
tioned, since an automatic heuristic detects no religious
marker in 62.7% of comments; a partial-input baseline
and per-target recall make the consequences visible.
• The Bangla T–V pronominal register is identified and
statistically validated as a correlate of the hate label,
under controls that separate a genuine sociolinguistic
signal from a corpus artefact.
• A six-item audit checklist is proposed for Bangla and
other low-resource hate speech resource papers.

TEDM-X: Transformer-Embedded Dirichlet Mixture Topic Discovery for Automatic and Interpretable Text Clustering

Conventional topic models rely on sparse word-count statistics and generally require the number of topics to be fixed before training. These assumptions weaken topic discovery when documents are semantically related but share limited surface vocabulary. This paper presents TEDM-X, a transformer-embedded topic discovery framework that combines Sentence-BERT document representations with a variational Bayesian Gaussian mixture model governed by a Dirichlet-process prior. Documents are cleaned, divided into overlapping sentence chunks, embedded with all-MiniLM-L6-v2, and L2-normalized before non-parametric mixture inference. The occupied mixture components define the effective topic set, while class-based TF–IDF produces interpretable topic keywords and labels. TEDM-X is evaluated against LDA, NMF, GSDMM, Top2Vec, and BERTopic on 20 Newsgroups and AG News using coherence, topic diversity, external clustering agreement, and internal cluster-validity measures. Relative to the strongest baseline, BERTopic, TEDM-X improves adjusted Rand index by 7.75% and 7.33%, normalized mutual information by 6.60% and 6.41%, and reduces the Davies–Bouldin index by 17.39% and 18.02% on the two datasets, respectively. The results indicate that combining contextual embeddings with adaptive Bayesian mixture inference yields more coherent, separated, diverse, and label-aligned topics without fixing the final number of active topics in advance.

GreenView: A Sustainability-Focused Deep Learning Pipeline with Explainable AI for Multiclass Plant Pathogen Classification

) Computational Energy Inefficiency: Training state-
of-the-art architectures from scratch consumes massive
computational power, contributing to a high carbon
footprint that contradicts the principles of sustainable
“Green AI.”
2) Resolution Discrepancies: Real-world plant pathogen
datasets are often heterogeneous, containing mixed im-
age resolutions (e.g., 256 × 256 and 512 × 512 pixels).
Standard resizing techniques often degrade fine-grained
visual features like necrotic lesion margins, causing
classification errors.
3) The “Black-Box” Problem: Standard deep learning
models lack interpretability. Without understanding why
a model classified a leaf as infected, farmers and
agronomists cannot fully trust its predictions.

Company-Specific Deep Learning Models for Next-Day Stock Price Prediction: A Comparative Study

The main contributions of this study are:

• We demonstrate that Encoder-only Transformer outper
forms GRU, CNN-LSTM and CNN-GRU, establishing a
performance baseline for the Bangladeshi capital market.

• We show that company-specific model selection yields
better results than applying a single architecture across
all stocks.

• We provide the first SHAP-based interpretability analysis
of deep learning stock prediction models on DSE data.

Beyond Balanced Benchmarks: Calibration, Uncertainty, and Deployment-Aware Evaluation of Prompt-Injection Detectors under Source Shift

This study examines prompt-injection detectors under source shift using calibration, frozen thresholds, and uncertainty-based deferral. It shows that better calibration does not always improve detection decisions. The results also reveal that deferral can still leave considerable attack leakage. Overall, the work provides a practical deployment-oriented evaluation using PromptShield and NotInject.

Multiclass Emotion Detection of Bangla Verse Text using NLP and Predictive Machine Learning Models

This study makes three main contributions. First, it introduces a new annotated dataset containing more than 18,000 Bangla poetic and song-lyric samples, categorized into seven emotion classes: Patriotism, Joy, Love, Sadness, Anger, Fear, and Surprise. The inclusion of Patriotism captures an important emotion-specific dimension of Bangla literary and cultural expression. Second, we develop a systematic modeling framework incorporating classical machine learning models, a hybrid stacking ensemble, and fine-tuned Bangla-BERT, enabling a comprehensive and consistent comparison of diverse modeling approaches. Third, we investigate the effects of class imbalance and provide detailed class-wise performance analysis, offering deeper insights into model strengths, limitations, and challenges in Bangla emotion classification.

A Physics-Grounded Explainable AI Framework for False Data Injection Attack Detection in Smart Grids

Smart grid transmission protection systems are increasingly vulnerable to False Data Injection Attacks (FDIAs), where manipulated relay measurements can compromise protection decisions while remaining difficult to identify using conventional data-driven methods. This paper proposes an explainable machine learning framework that combines reliable FDIA detection with transparent decision interpretation for transmission protection applications. The proposed pipeline employs consensus feature selection by combining Information Gain, Random Forest, and Support Vector Machine rankings, with Correlation based Feature Selection applied as an independent confirmatory check, followed by the best performing Extra Trees configuration identified through cross validated model comparison under a strictly separated training and testing protocol. To improve interpretability, SHAP, LIME, permutation importance, and DiCE counterfactual analysis are jointly incorporated, while a physics consistency audit evaluates whether the dominant relay measurements exhibit electrically plausible behavior. Experiments were conducted on the Mississippi State University and Oak Ridge National Laboratory transmission protection dataset, containing 78,377 labeled observations across fifteen cyber physical scenarios. The final model achieved 93.65% mean accuracy, 0.8848 F1-score, 0.9327 precision, 0.8416 recall, and 0.9820 ROC-AUC, with stable performance across five independent data splits. Unlike conventional black box FDIA detectors, the proposed framework provides global and local explanations together with relay oriented physical validation, enabling operators to understand why an alarm is generated and whether the decision is consistent with underlying electrical behavior. These findings support the development of trustworthy, operationally interpretable protection systems for future smart grids.

DRAFT

Green Guard: A Convolutional Neural Network Based Desktop Application for Automated Crop Leaf Disease Detection and Advisory

Green Guard bridges CNN-based disease detection with practical agricultural decision support through an integrated desktop solution for three major crops.

A Simulation-Based Machine Learning Framework for Quantum Noise-Channel Classification

Quantum teleportation (QT) is the most essential component of quantum communication, but its performance falls significantly when noisy intermediate-scale quantum (NISQ) channels are present. This paper presents a physics-informed machine learning (ML) framework for classifying quantum noise channels in teleportation systems. We employ Kraus operators to model five representative noise processes: depolarizing, amplitude damping, phase damping, bit flip, and phase flip. Additionally, we simulate their effects on teleportation fidelity and entanglement. A dataset comprising 15,360 instances is generated using the Monte Carlo method, featuring 35-dimensional vectors derived from quantum information metrics, including fidelity, entropy, purity, and Quantum Fisher Information (QFI). Multiple classifiers are evaluated, including Support Vector Machines (SVM), Random Forest (RF), Multilayer Perceptron (MLP), and Gradient Boosting (GB). The results indicate that ensemble methods demonstrate superior performance compared to alternative models. Specifically, GB attained the highest accuracy of 96.5%, ensuring strong per-class performance across all categories of errors. These findings highlight the potential of physics-informed ML for developing noise-aware and reliable quantum communication systems.

Performance Analysis of Frequency Governor Controllers under Implementing Free Governor Mode Operation in the Bangladesh National Grid

Grid frequency is a critical parameter that must
be maintained within an acceptable range to ensure the re-
liability, stability, and economic operation of modern power
systems. Free Governor Mode Operation (FGMO) has gained
significant importance in Bangladesh for providing primary
frequency response and enhancing grid resilience under varying
load conditions. This paper presents a performance analysis of
frequency governor controllers operating under FGMO using
operational data collected from Chandpur Power Generation
Limited (CPGL). The study investigates governor droop charac-
teristics, frequency regulation mechanisms, and the operational
differences between FGMO and conventional fixed power mode.
A mathematical framework based on droop control is developed
to describe the relationship between frequency deviation and
active power response. Experimental results obtained from a
one-hour FGMO field test demonstrate stable system operation
with a 4% governor droop setting and an 85% nominal engine
setpoint, maintaining the grid frequency within 49.54–50.58 Hz
while producing 60.48 MWh of energy with a net heat rate
of 8078.29 kJ/kWh. Comparative analysis shows that although
fixed power mode offers slightly higher fuel efficiency under
high loading conditions, FGMO significantly improves frequency
regulation, load sharing, operational reliability, and overall grid
stability. The results indicate that widespread implementation
of FGMO can reduce frequency excursions, minimize blackout
risks, facilitate the reliable integration of future renewable energy
resources, and support compliance with international primary
frequency response standards in the Bangladesh National Grid.
Index Terms—Bangladesh National Grid, Free Governor Mode
Operation (FGMO), Frequency Control, Governor Droop Con-
trol, Grid Stability, Primary Frequency Response.

An Optimizing MAC Protocol for WBANs Managing Multi-class and Multi-load in IEEE 802.15.6 Superframes Using Hysteretic Q-learning

Wireless Body Area Networks (WBANs) have become a fundamental technology for continuous healthcare monitoring by enabling reliable communication among wearable and implantable biomedical sensors. Since WBAN applications generate heterogeneous traffic with different Quality of Service (QoS) requirements, efficient Medium Access Control (MAC) protocols are essential to provide reliable data delivery while maintaining low latency and energy consumption. Although the IEEE 802.15.6 standard supports multiple access phases for different traffic priorities, its predefined superframe structure cannot efficiently adapt to dynamically varying traffic loads, leading to underutilized bandwidth, increased packet collisions, and reduced network performance. To address these limitations, this paper proposes an Optimizing Multi-Class Multi-Load Handling MAC (OMMH-MAC) protocol that integrates Hysteretic Q-learning (HQL) into the IEEE 802.15.6 superframe for adaptive slot allocation and dynamic traffic scheduling. The proposed protocol first classifies network traffic and estimates network load. Based on the learned HQL policy, the Body Coordinator (BC) dynamically allocates TDMA slots within the IEEE 802.15.6 superframe. Simulation outcomes exhibit that the proposed protocol achieves faster convergence, higher throughput, lower packet collisions, and reduced energy consumption and making it suitable for dynamic healthcare monitoring applications.

External Validation of a Three-Backbone Heterogeneous CNN Ensemble for Six-Class Skin Lesion Classification: A Cross-Dataset Study

This paper addresses a gap in automated skin-lesion classification: models are almost universally evaluated in-domain, on a held-out split of the dataset they were trained on, and a survey of eleven recent studies finds essentially none that apply external cross-validation on images from an independent acquisition site. We therefore adopt a strict protocol — trained only on HAM10000, tested only on the external Derm7pt dermoscopic dataset over the six lesion classes common to both, with no target-domain fine-tuning and every design choice fixed a priori — and show that a widely cited InceptionV3+DenseNet121 weighted-fusion design reported at 92.27% on a held-out split of its own training set operates at 70.89 ± 0.66% under this regime, demonstrating that published headline accuracies carry little information about behaviour at a new site. Within this protocol we propose a three-backbone heterogeneous ensemble that adds an architecturally distinct, frozen-base ConvNeXt-Large probability stream to that fusion, and show it improves every macro metric on the identical 963-image external test set across three seeds (accuracy 73.38 ± 2.25%, macro-F1 56.71 ± 3.05%, macro-AUC 90.38 ± 0.90%; +2.49, +4.41 and +3.10 respectively), with the clearest benefit on basal cell carcinoma — a malignancy with only 42 external test images, precisely the low-support regime in which a single backbone is least reliable — while no individual stream matches the fusion, indicating that the three backbones make complementary rather than redundant errors. The significance is thus both methodological and architectural: it places a widely used fusion design on an honest external footing and shows that architectural heterogeneity, rather than added depth or attention, is what survives the transfer; we report the result with its bounds stated explicitly, since the comparison is at the configuration level (the three-backbone run also changes augmentation and batch size) over three seeds on a single external dataset.

AI-Driven Transboundary Stacking Framework for Flash Flood Risk Prediction in the Sylhet Basin, Bangladesh

This study contributes an AI-driven transboundary stacking framework for flash flood risk prediction in the Sylhet Basin. It integrates local rainfall with lagged and cumulative upstream rainfall from Meghalaya and Assam using Random Forest, Gradient Boosting, and Logistic Regression. The proposed framework achieved an F1-score of 0.9855 and ROC-AUC of 1.0000, demonstrating strong potential for early flood-risk detection.

RSMS: A FreeRTOS-Based Priority-Scheduled Architecture for Real-Time Multi-Hazard Detection Using Queue-Driven Task Communication and Bluetooth Alerting

This paper presents RSMS, a novel Real-Time Operating System (RTOS)-based safety monitoring architecture that introduces three key contributions to the field of embedded safety systems. First, it establishes a three-tier priority-scheduled framework on an ESP32 microcontroller running FreeRTOS, where four independent sensing tasks (flame, gas, IR obstacle, and Bluetooth SOS) are assigned differentiated priorities and pinned across two processor cores, ensuring that life-threatening hazards such as fire are serviced with deterministic preemption over lower-priority events. Second, it implements a decoupled producer-consumer model through a single FreeRTOS queue, which separates hazard detection from alert response, enabling modular expansion without modifying core logic and preventing resource conflicts during concurrent event handling. Third, the system integrates practical engineering solutions including a five-cycle debounce filter for IR sensors to suppress false positives, a hysteresis-based threshold for gas detection to prevent alert oscillation, and a Bluetooth-based emergency SOS channel—all realized at a total hardware cost of USD 4.23, making it accessible for resource-constrained deployments. Collectively, these contributions advance the state of the art by demonstrating that RTOS-based priority scheduling, when combined with structured inter-task communication, can reliably handle concurrent multi-hazard scenarios that conventional single-loop systems cannot guarantee, as validated through four independent serial and Bluetooth-terminal capture sessions.

A Rollback-Resistant Negotiation Protocol for Context-Adaptive Post-Quantum Key Encapsulation

The major
contributions of this work are summarized as follows:
• A downgrade (rollback) attack is identified and demon-
strated against naive context-adaptive Kyber key encap-
sulation, in which an active on-path adversary forces
negotiation to a weaker security level without detection
by either endpoint.
• A rollback-resistant negotiation protocol is proposed that
binds the negotiated security level into a transcript hash
and verifies it via a shared-secret-derived MAC, causing
any tampering to result in handshake rejection.
• A comprehensive experimental evaluation is conducted,
implementing both the vulnerable baseline and the pro-
tected protocol across all three Kyber security levels, to
demonstrate the attack and its mitigation and to quan-
tify the overhead introduced by the proposed protection
mechanism.

Deep Sense: Deep Learning for Early Staging the Onset of Diabetes and Optimizing Insulin Dosages Emanating Recognition of Activity Patterns

The main research contribution of Deep Sense is a smartphone-based, low-power framework that combines 14-class diabetes-relevant Human Activity Recognition using LSTM, GAN-based synthetic sensor-data augmentation, and cosine-similarity-based diabetes risk scoring against a clinically sourced diabetic-patient dataset. The approach achieved 98.48% test accuracy, while the risk-scoring stage produced a 57.39% similarity score that was corroborated by an A1C result of 6.1% in the evaluated subject.

Dual-Flow Knowledge Distillation with Category-Adaptive Inference for Unsupervised Industrial Anomaly Detection

Our contributions are:
1. We propose a dual-branch scoring architecture jointly training hierarchical distillation and per-level normalizing flows on a shared frozen teacher pyramid, fused via depth-weighted, standardized combination.
2. We include frequency-Aware Attention (FAA) residual block for the student, re-weighting channel responses via learned pooling over grouped frequency-oriented projections, with a cosine+MSE distillation loss.
3. We also introduce category-adaptive policy (selective augmentation, adaptive top-k scoring, soft foreground masking, optional rotation TTA) from simple, interpretable object priors rather than exhaustive tuning.
4. We perform transparent ablation showing that two of four fixes did not generalize and were reverted; and results on all 15 MVTec AD categories: mean image-AUROC 94.07%, mean pixel-AUROC 96.71% are achieved, which can be considered a competitive score.

An Attention-Enhanced Lightweight CNN with Cross-Domain Robustness for Edge-Based Plant Disease Identification

The main contributions of this paper are threefold. First, this
work integrates CBAM into a MobileNetV3-Large backbone
to improve attention to plant disease lesions while keeping the
model suitable for edge-oriented deployment. The resulting
architecture contains only 4.36 million trainable parameters,
which makes it more practical for low-resource agricultural
monitoring than heavier CNN backbones. Second, this work
uses Grad-CAM to examine whether the model is learning
visually meaningful disease regions. The attention maps show
that the classifier focuses on lesions and infected tissue rather
than relying only on background texture or leaf placement.
Third, this work evaluates the same trained checkpoint on
both PlantVillage and PlantDoc, reporting both controlled
laboratory accuracy and the cross-domain failure that appears
when the model is tested on field-style images.

An Analysis of the Financial Performance of National Credit and Commerce Bank PLC (NCC Bank PLC)

This study contributes a comprehensive eleven-year financial performance analysis of NCC Bank PLC from 2015 to 2025, integrating profitability, asset quality, liquidity, and capital adequacy indicators. Unlike a single-year assessment, the study identifies long-term trends and changes in the bank’s financial condition. A key contribution is the identification of the bank’s profitability pressure during 2018–2024 and its significant recovery in 2025. The study also links this recovery with improved asset quality, particularly the decline in the non-performing loan ratio to 4.12%, along with strong liquidity and capital adequacy. Furthermore, the study highlights an interesting market-valuation finding: although EPS nearly doubled to Tk. 4.38 in 2025, the P/E ratio declined to 2.79 times. Overall, the study provides an integrated view of NCC Bank’s financial stability, performance recovery, and investor-related performance.

Automated EEG-Based Schizophrenia Detection by a Feature Extraction and Machine Learning Algorithm

This study harmonizes two distinct public datasets (IBIB PAN and ASZED, encompassing 181 subjects and over 60,000 epochs) into a standardized 19-channel framework and engineers a unified tri-domain feature space fusing Power Spectral Density (PSD), Functional Connectivity (coherence and PLV), and Continuous Wavelet Transform (CWT) dynamics. A systematic evaluation across twelve feature-model combinations demonstrates that classical ensemble classifiers (Random Forest and XGBoost) achieve 98.7% accuracy and a 0.99 F1-score, confirming that lightweight, interpretable machine learning can match complex deep neural architectures for non-invasive schizophrenia screening.

Vision-Based Traffic Signal Control with an Enhanced Deep Q-Network: Evaluation and Reward Decomposition on a Camera-Derived Benchmark

An end-to-end vision-based controller. We present a dueling, double-target EDQN with 9.54 million parameters that maps four raw 128×128 RGB camera views directly to a signal phase, removing the detection and state-estimation stages that intervene in conventional pipelines.
A reproducible offline protocol. We define a complete ten-stage pipeline – preprocessing, MDP construction, training, greedy evaluation, metric estimation and baseline comparison – over a camera-derived corpus of 728 synchronised four-view frames, with every reported figure traceable to a stated closed-form estimator.
Quantified performance against baselines. The EDQN attains a cumulative return of -4358.4 against -9748.4 for a random policy and -9868.4 for a fixed-time policy, a 2.24× improvement, and identifies the highest-demand approach in 99.0% of steps compared with 25.0% and 23.3% respectively.
A reward decomposition diagnostic. We derive a closed-form decomposition of the reward and of the delay estimator that attributes the entire performance gap to the approach-selection term, proves that queue, waiting time, throughput and speed are policy-invariant under open-loop replay, and shows that an apparent 0.47 s delay regression is an estimator artefact rather than a behavioural effect.

Integrated Solar Tracking and Active Water Cooling for Enhanced Photovoltaic Performance: Experimental Validation and PVsyst-Based Techno-Economic Assessment

Fixed-tilt photovoltaic (PV) panels lose a significant
portion of the solar resource since they are not typically tilted 90 degrees to the sun’s rays, and the amount of solar energy they receive that is not usable because it is too hot increases cell temperature, which also reduces PV efficiency. For that reason this paper represents the design, construction and controlled experimental testing of an integrated system, which is a combination of light dependent-resistor (LDR) solar tracker mounted on a single axis and automatic active water cooling loop that is engaged at a maximum allowable panel temperature of 40°C and released
at 25°C under the supervision of an Arduino Nano controller is reported where the cooling system was automatically activated when the PV module temperature reached 40 ◦C and deactivated
when the temperature decreased to 25 ◦C. Three operating configurations such as fixed-tilt, single-axis tracking, and tracking with active cooling were experimentally evaluated under identical outdoor conditions during a one-month measurement campaign
from 09:00 to 18:00. The results demonstrated a consistent improvement in PV performance across the investigated configurations. The mean electrical efficiency increased from 2.86% for the fixed panel to 3.80% with single-axis tracking and 4.19% with integrated tracking and active cooling. Compared with the fixed configuration, single-axis tracking and integrated tracking with cooling achieved power gains of 32.5% and approximately 55.6%, respectively. Active cooling reduced the maximum module temperature from 49.7 ◦C to approximately 40 ◦C and provided an additional 17.3% performance improvement compared
with the tracking-only configuration. Regression analysis yielded R2 > 0.80 for the fixed and tracking with having cooling configurations, indicating the importance of thermal regulation in maintaining a consistent temperature efficiency. To evaluate
system level feasibility, a grid-connected 1 kWp PV system was additionally modeled using PVsyst for the study location. The simulation predicted an annual energy yield of 1499.2 kWh, an average performance ratio of 76.2%. Overall, the framework demonstrate that coordinated solar tracking and threshold-based
active cooling can provide complementary optical and thermal benefits, while the PVsyst assessment indicates promising technoeconomic feasibility for distributed PV deployment in high temperature and high-irradiance environments that provides a baseline for the development of more advanced adaptive tracking
and predictive thermal-management strategies for photovoltaic systems.

Blockchain-Assisted Temporal Graph Learning for Trustworthy Supply Chain Forecasting: A Bangladesh Case Study

1)A blockchain-backed provenance mechanism for temporal
supply-chain data, using an off-chain-data / on-chain-hash
design with a minimal audited Solidity contract;
2) Integration of verified temporal data with graph-based
production forecasting (persistence, LSTM, and GCon
vLSTM) under a strictly chronological protocol;
3)A controlled data-tampering experiment that quantifies
how tampering degrades forecasting; and
4) Quantification of tamper detection, verification-driven re
covery, and the provenance-layer overhead.

AgroConv-Eff: A Hybrid ConvNeXt–EfficientNetV2 Architecture with Adaptive Feature Gating for Multi-Crop Plant Disease Detection

Timely diagnosis of crop disease directly shapes yield outcomes, and for many smallholder farmers, a trained eye scanning a leaf is still the only diagnostic tool available. This paper presents AgroConv-Eff, a hybrid deep learning architecture that fuses ConvNeXt-Tiny and EfficientNetV2-S feature-extraction pathways through an adaptive channel-wise gating mechanism for automated multi-crop plant disease classification. The system was trained and evaluated on a 38-class, 14-crop plant disease dataset derived from the PlantVillage image collection, comprising leaf images spanning multiple crops and disease categories. Using ImageNet-pretrained backbones and transfer learning, the proposed model achieved 99.34% test accuracy and a macro-F1 score of 98.77%. To rigorously position this result, we additionally trained and evaluated three strong baselines under an identical protocol: MobileNetV2 (98.73% accuracy), ConvNeXt-Tiny alone (99.37% accuracy), and EfficientNetV2-S alone (99.16% accuracy). The comparison shows that, on this dataset, the single-backbone ConvNeXt-Tiny model is competitive with and marginally exceeds the proposed hybrid in raw accuracy while using substantially fewer parameters and lower inference latency, a finding we discuss openly along with per-class error analysis. To isolate the specific contribution of the proposed adaptive gating mechanism, we further design a controlled ablation comparing gated fusion against simple concatenation and fixed-weight averaging under identical backbones and training settings. Overall, this work contributes a fully reproducible training and evaluation pipeline, a systematic four-model comparison under one identical protocol, an ablation protocol isolating the fusion mechanism’s contribution, and an honest efficiency-accuracy trade-off analysis with explicitly stated limitations for multi-crop plant disease detection.

Exploiting Federated Learning Approach for Efficient Bandwidth Utilization in IoT-Based Surveillance System

As Internet of Things (IoT)-based surveillance devices expand, maintaining data privacy and bandwidth efficiency has emerged as a crucial concern. Security vulnerabilities, bandwidth congestion, and latency are common problems with traditional cloud systems. This paper uses edge computing, YOLOv9 object detection, and H.265 video encoding in a surveillance architecture based on federated learning (FL) to address these problems. The proposed system retains and compresses only motion-based frames before transmitting them to the central server, considerably minimizing duplicate data transfer. The FL model is collaboratively trained across three edge clients without sharing raw video, thereby preserving privacy. After 200 rounds of communication, the global model achieved a mean Average Precision (mAP@50) of 94.6%, a precision of 91.5%, a recall of 88.4%, and an F1-score of 89.6%. Furthermore, the system reduced an experimental video size from 32 MB to 2 MB and duration from 11 seconds to 5 seconds by filtering out static frames. Additionally, field experiments on several video footages confirmed the robustness of the proposed technique. For page limitations, only one sample video result is described in detail in the main text and summarizes the performance over all tested videos in a comparison table. These results validate that the proposed combination of FL, intelligent frame filtering, and H.265 compression leads to a scalable, bandwidth-efficient, and privacy-preserving IoT surveillance.

AnemiaScreen: Explainable Fusion of Handcrafted and Deep Features for Non-Invasive Anemia Screening from Palpebral Conjunctiva Images

We develop a hybrid screening pipeline that fuses an in-
terpretable 212-dimensional conjunctival descriptor with
the embedding of a fine-tuned Swin-T backbone, cover-
ing conjunctiva localisation, illumination normalisation,
fusion and a decision threshold fixed before testing.
• We provide a systematic analysis of model performance
in a small-data, single-cohort setting by combining ab-
lation experiments with learning-curve analysis, allowing
us to identify which components and how much training
data contribute to the observed predictive performance.
• We build and deploy AnemiaScreen, an Android applica-
tion that runs the pipeline fully offline, stores screening
records on the device and states plainly where it falls
short of the reported model.

A Probabilistic Consistency and Profile-Based Framework for Protein Multiple Sequence Alignment

Multiple Sequence Alignment (MSA) is a crucial bioinformatics task for locating conserved regions, evolutionary relationships, and functional similarities among biological sequences. The accurate alignment of proteins continues to be difficult because of factors such as sequence divergence, insertions and deletions, the positioning of gaps, and the extensive search space involved in the alignment process. A probabilistic consistency and profile-based framework for protein MSA is introduced in this study. It integrates pair-HMM posterior probabilities, posterior consistency, guide-tree construction, expected-accuracy profile alignment, and progressive profile merging. Initially, the framework builds a preliminary alignment via progressive profile alignment and assesses residue correspondences based on posterior probabilities. The proposed method was assessed on the BAliBASE RV11 and RV12 benchmarks using Sum-of-Pairs (SP) and Total-Column (TC) scores with FastSP. It reached average SP/TC scores of 0.6751/0.3644 on RV11 and 0.9096/0.6794 on RV12, with mean alignment runtimes of 1.44 and 1.87 seconds, respectively. The outcomes show a competitive accuracy for pairwise alignment while also highlighting chances for improving complete-column recovery.

Merit-Order Economics of Free Governor Mode Operation: A Business-Analysis Framework for Heat-Rate-Based Dispatch Priority and Frequency-Regulation Value in the Bangladesh National Grid

Although Free Governor Mode Operation (FGMO) has been extensively studied technically (droop response, frequency deviation, plant fuel efficiency, etc.), the business case for what plant companies are called into FGMO duty, and the economic logic they are based on, has not been investigated. This omission is important because the dispatch decision, which is taken by the system operator like the National Load Despatch Centre (NLDC) of Bangladesh, represents a decision on the allocation of resources with tangible impacts on generator revenue, regulation, and the welfare of industrial customers. This paper builds on the operations and energy-business literature by providing a business-analysis framework based on the engineering parameter, generator heat rate, as a proxy for marginal cost, and its connection to a firm-level economic decision—that of NLDC’s merit-order dispatch practice. We illustrate how lower cost generating units are gradually relegated to higher and higher dispatch levels, that is, they perform both base-load and primary reliability service, but are paid the same rates for both services. We then explore the business case for such an agreement based on Bangladesh-specific estimates of the value of lost load and the
cost of unplanned industrial outages, and conclude that the cost of the fuel premium that the frequently dispatched units pass on to generators is, in most realistic scenarios, well outstripped by the value of the reliability that the unit provides to the rest of the economy. The implications from these findings are clear for
management and regulation: generation companies are currently claiming a reliable-service cost, but without formal recognition, causing a mismatch between private incentives and system-wide value creation. To fix this, the paper suggests a compensation scheme based on heat rate that regulators can implement in
Bangladesh, for instance, the Bangladesh Energy Regulatory Commission (BERC), to create incentives for the firms that are aligned with the outcomes of providing grid reliability, and which could be transferable to other developing electricity markets.

A Generalized Unit-Cell-Based Multilevel Inverter with Reduced Components, Low THD, and Scalable Output Voltage Levels

Multilevel inverters (MLIs) have been widely employed in modern power conversion systems because of their enhanced output voltage quality, reduced voltage stress, and lower total harmonic distortion (THD). This paper presents a generalized unit-cell-based multi-source multilevel inverter topology capable of generating scalable output voltage levels through the cascading of identical unit cells. The proposed topology is developed and evaluated for 15-level and 31-level configurations using only 8 switches with 3 DC sources and 10 switches with 4 DC sources, respectively. Optimum Nearest Level Modulation (ONLM) is employed for generating the required switching pulses. The proposed configurations are validated through MATLAB/Simulink simulations under resistive, RL, and dynamic load conditions. The obtained voltage THDs are 5.39% and 4.71% for the 15-level and 31-level configurations, respectively. The results demonstrate that the proposed topology offers reduced component requirements, low harmonic distortion, and flexible scalability, making it a promising solution for high-quality power conversion applications.

Robust Ransomware Detection Through Cross-Family Evaluation and Behavioral Feature Analysis

The main contributions of this work are summarized as follows:

Cross-family evaluation: We evaluate ransomware de
tection on unseen families using cross-family and 12-fold
LOFO protocols.
Comprehensive evaluation: We compare seven classi
fiers using Accuracy, Precision, Recall, F1-score, Speci
ficity, and AUC.
Feature ablation: We analyze static and dynamic fea
tures to identify which feature type provides better gen
eralization across unseen families.
Ensemble analysis: We evaluate voting and stacking en
sembles against individual classifiers under cross-family
and LOFO settings.
Explainability: We identify the most important transfer
able features using permutation importance on held-out
families.
Category-level validation: We perform five-way classifi
cation to verify ransomware-specific detection capability.

Fruits Classification Using a Custom Dataset and Ensemble CNN Models with Real-Time Web-Based Deployment

Abstract—The advancement of deep learning has transformed
many fields, including agriculture, where automation and intelligent systems are becoming increasingly important. One
important task is fruit sorting and identification. In this study,
we developed a fruit classification system using Convolutional
Neural Network (CNN) models. Instead of relying on an existing
dataset, we created our own dataset to better represent real
world conditions. A total of 3,200 images were collected across 32
fruit classes under different lighting, backgrounds, and camera
angles. The dataset includes both commonly consumed fruits
and regionally important varieties such as bael, monkey jack,
and wood apple that are not present in most public fruit
datasets, making it more regionally relevant. Before training,
the dataset was carefully prepared by fixing rotation issues,
removing duplicate and corrupted files, and resizing all images
to a standard format. We also applied simple data augmentation
techniques such as flipping and brightness adjustment to help
the models learn better. Several CNN architectures were tested,
including EfficientNet, MobileNet, DenseNet, ResNet50, VGG19,
InceptionV3, and Xception.

Memory Fossils in Large Language Models: A Mechanistic Framework for Quantifying Residual Latent Knowledge Post-Unlearning

The concept of the Memory Fossils framework is presented to discover any potential leftover information existing in LLMs after implementing machine unlearning. The framework distinguishes three types of residues: Activation Fossils, Attention Fossils, and Gradient Fossils. It also proposes deriving a score called Fossil Score by combining different approaches, such as probing, causal reconstruction, and output deviations. The results indicate a low output recall does not guarantee complete suppression of knowledge in question, indicating a need to examine internal representations along with behavior.

A Low-Cost Embedded Data Acquisition System For Concrete RCP Test

The durability of reinforced concrete structures is strongly governed by chloride ion penetration, which accelerates corrosion of embedded reinforcing steel and shortens service life. The Rapid Chloride Permeability Test (RCPT) is a widely adopted method for evaluating concrete’s chloride resistance, involving a 60V DC voltage applied across a specimen with current monitored over six hours to determine total charge passed. However, commercially available RCPT systems with automated data acquisition remain costly, limiting accessibility in resource-limited contexts. This study addresses this gap through a low-cost, microcontroller-based embedded data acquisition system for RCPT applications, integrating voltage and current sensors, an LCD display, and an ESP32-S2 microcontroller for real-time monitoring, control, and data storage. The system logs voltage and current at defined intervals, computes total charge passed, and stores results for post-test analysis, with calibration against digital multimeters ensuring measurement accuracy. Results show the system delivers precise, automated performance comparable to commercial equipment at substantially lower cost, while addressing key challenges including noise reduction, calibration accuracy, and long-term reliability. The proposed system offers a practical, affordable alternative for chloride permeability testing, particularly for laboratories in resource-limited settings.

Keywords—Rapid Chloride Permeability Test, RCPT, Concrete Durability, Chloride Permeability, ESP32-S2, ADS1115, voltage sensor, current sensor, Embedded System, Data Acquisition System (DAS).

Runoff Forecasting and Flood Risk Classification in Sunamganj: A Comparative Analysis of Statistical, Machine Learning, and Deep Learning Models

This study has developed an near real-time runoff forecasting and flood risk classification framework for Sunamganj. Statistical, Machine Learning and Deep Learning models such as Persistence, Linear Regression, Random Forest, XGBoost, LSTM and GRU have been used to predict 1–24 hours of runoff and compare models. Hydrological patterns have been analyzed using time, lag and rainfall-runoff related features. In addition, three flood risk classes—Normal, Moderate and High—have been created based on runoff and the classification performance has been evaluated. This study provides an effective flood early warning approach for flood-prone areas of Bangladesh.

Beyond Point Accuracy: Regime-Conditional SHAP and Walk-Forward Validation for Australian GDP Growth Forecasting

Our contributions are:
• A walk-forward (rolling-origin) evaluation of six forecasting models spanning classical econometrics, ensemble ML, and quantile-regression deep learning, with bootstrap confidence intervals and formal Diebold-Mariano
significance testing between every model pair.

• Aregime-conditional SHAP analysis that quantifies how each model’s reliance on macroeconomic features shifts between crisis and calm periods—an interpretability angle largely absent from existing GDP forecasting litera
ture.

• A learning-curve analysis quantifying the training-data threshold at which each model class stabilizes, directly addressing the small-sample concern rather than treating it as an unexamined limitation.

SyllabusSherlock: A Curriculum-Aware and Knowledge Graph-Enhanced Generative AI Tutoring System for NCTB Education

SyllabusSherlock is a curriculum-aware AI tutoring system built around the NCTB curriculum. At its core, it links a Knowledge Graph of textbook topics and prerequisites with textbook-based retrieval and an estimate of what the learner already knows. Together, these let the system move beyond simple question-answering, offering prerequisite-based learning paths, adaptive assessment, and practice tailored to each student. We also present an initial evaluation of the generated tutoring responses, looking at how correct, textbook-grounded, relevant, clear, and personalized they are.

AgriSemNet: A Semantic-Aware Hybrid Deep Learning and Knowledge Graph Architecture for Precision Crop Recommendation

Most crop recommendation systems either act as uninterpretable black boxes or struggle to process raw sensor data. In this paper, we built AgriSemNet to bridge this gap by combining a tabular MLP with a Graph Attention Network (GAT) and an external 42-node Agronomic Knowledge Graph. Instead of only predicting a single static crop label, our system generates a practical 4-tier recommendation: the primary crop, a low-risk backup crop, a sustainable crop rotation sequence, and clear biological justifications for the farmer. We also built in two-stage soil safety guardrails to flag extreme or toxic soil conditions before planting. On a benchmark 22-crop dataset, our model achieves 98.79% accuracy and stays remarkably stable against IoT sensor noise, dropping only 1.52% under 20% Gaussian noise, where standard models like Random Forest and XGBoost degrade significantly.

Handwritten Prescription Recognition Through Transfer Learning Models Evaluation Under Multiple Dataset Partitions

1. Before the model evaluation, a duplicate-aware dataset verification and leakage-control framework was developed with the help of exact image hashing and perceptual grouping.

2. Four different train-validation-test partition strategies were systematically evaluated to investigate the effect of data partitioning on model performance.

3. The five transfer-learning architectures including VGG16, DenseNet121, ResNet50, EfficientNet-B0 and MobileNetV3-Large are trained on the same training and evaluation settings.

4. To prevent test-set selection of the final model, a validation-based model-selection strategy was used, with Macro-F1 as the validation measure.

5. To test the stability and robustness of the selected configuration, and to avoid visual-group leakage between folds, five-fold stratified cross-validation is employed.

Synthetic Data Augmentation for Bangla Hate Speech Detection Using Slot-Based Semantic Extraction

Novel Framework for Bangla Hate Speech: Developed a robust multi-class classification framework (Hate, Non-Hate, Sarcastic) tailored for implicit hate and sarcasm in low-resource Bangla.

Slot-Based Semantic Extraction: Introduced a structured slot extraction method (TARGET, Negative Expression, Neutral Context, Sarcastic Cue) to capture contextual semantic roles.

Balanced Synthetic Data Augmentation: Built a template-driven, slot-replacement synthetic augmentation technique that expanded a 5,000-comment corpus into a fully balanced 16,302-comment dataset.

State-of-the-Art Performance: Achieved 91.90% accuracy and 91.76% Macro F1-score using BanglaBERT, along with a lightweight TF-IDF+LR baseline reaching 85.00% accuracy, outperforming existing benchmarks.

Three-Layer Forensic Fingerprinting for Insider Leak Attribution: Robust Tracing, Tamper Localization, and Content Binding

To address the challenge of insider leak attribution, this paper presents a three-layer DCT-domain forensic fingerprinting architecture that jointly provides: (i) robust user tracing via QIM-embedded fingerprints with Reed-Solomon error correction; (ii) key-shared tamper detection with block-level localization; and (iii) cryptographic content binding that rejects transplant attempts. All three layers are embedded in a single pass over disjoint DCT bands to prevent interference. Evaluated on 138 benchmark images, the scheme identifies the correct leaker in every benign test, flags every tested manipulation with 0.46-0.88 IoU, completely blocks transplant attempts, and identifies colluders in averaging attacks with zero false positives. Finally, a forced trade-off analysis demonstrates that erasing the fingerprint necessarily destroys the content (dropping quality to at most 25.2 dB PSNR) or triggers a full tamper alarm, proving that no clean, untraceable copy can be laundered.

A Deep Learning Framework for Cross-Medium Deepfake Detection in Independent Images

The rapid progress in deep generative models has led to the creation of hyper-realistic synthetic media that easily evades human perception. While modern networks can generate convincing fake assets, existing forensic tools struggle to generalize because they are optimized either for isolated images or continuous video streams, but rarely both. To address this limitation, we propose a unified, cross-medium deep learning framework designed to evaluate authenticity across independent images. The proposed framework evaluates independent images by utilizing a custom Convolutional Neural Network (CNN) baseline alongside advanced Vision Transformers (ViT) and Hierarchical Mixed-Attention (MaxViT) architectures to capture micro-texture spatial anomalies. The framework was comprehensively evaluated on large-scale datasets under real-world data distributions. In the image-level analysis, the baseline CNN achieved an accuracy of 55.74%, which improved to 89.11% with the ViT backbone, and peaked at 93.40% using the MaxViT architecture. Ultimately, this work provides a highly generalizable and robust solution that significantly advances digital media forensics and multimedia verification systems.

Leakage-Aware Ordinal Prequential Forecasting of District-Level Conflict Severity in Bangladesh

Abstract—Reliable district-level conflict forecasting requires
models that distinguish ordered severity while controlling temporal
leakage, geographic heterogeneity, class imbalance, reporting
processes, and non-stationarity. We propose a leakage-aware
prequential framework for forecasting next-month Low, Medium,
and High conflict severity across all 64 districts of Bangladesh.
The study integrates three data sources: (1) ACLED, comprising
31,613 reconciled conflict-event records; (2) NASA POWER,
providing precipitation, temperature, and wind variables; and (3)
geoBoundaries ADM2, providing district-level spatial information.
Their integration produces a complete 11,904-observation districtmonth panel spanning 186 months. Severity is defined by a
study-specific analytical index combining next-month events and
reported fatalities, with the High cutoff (c = 6) frozen from an
initial 60-month history.

Concept Drift Detection in Streaming Bangla Text Using an Ensemble Statistical Framework

1. We propose an ensemble that combines a BanglaBERT-plus-autoencoder reconstruction-error signal with a label-distribution signal through weighted voting over three statistical detectors.
2. We provide a controlled benchmark against six established detectors (ADWIN, DDM, EDDM, HDDM-W, Page-Hinkley, and KSWIN), reporting F1-Score and false-alarm rate on each drift type.
3. Through a per-detector analysis, we characterise which signals contribute to detection, showing that the label-distribution channel carries most of the discriminative power in our dataset.

Explainable Federated Early-Warning for Commercial Banks: Privacy-Preserving Financial-Distress Prediction and Decision Support in Bangladesh

Bank supervisors need early-warning models that learn from multiple institutions without centralizing confidential bank records. This paper evaluates an explainable cross-silo federated learning framework for one-year-ahead bank-distress prediction in Bangladesh. The compiled panel contains 780 bankyear records for 57 banks over 1997–2025; after forward target construction, 614 observations remain for 2013-2024, with 452 training and 162 strictly out-of-time test observations. Distress
is defined as a next-year breach of either a 10% NPL or 10% CAR threshold. Seventeen CAMELS-derived predictors are winsorized and standardized, and five unbalanced silos train classweighted logistic models aggregated by sample-weighted FedAvg. Centralized Logistic Regression, Random Forest, XGBoost, and Explainable Boosting Machine provide reference baselines, while FedProx, clustered personalization, and differentially private FedAvg test robustness and privacy. FedAvg reaches ROC-AUC 0.9685 versus 0.9714 centrally, retaining 99.7% of reference discrimination without moving raw records. At an RDP-accounted operating point of ϵ ≈ 3.52, δ = 10−5
, DP-FedAvg reaches ROCAUC 0.926. SHAP identifies NPL ratio, ROE, log Z-score, ROA,
and loan-to-assets as leading drivers; independent CAMELS ablation confirms asset quality, earnings, and capital as the most consequential groups. A two-year horizon loses substantial threshold-dependent performance, and a supervisory watchlist
translates probabilities into risk bands and drivers. The results support privacy-aware collaborative bank-risk screening while clearly distinguishing retrospective federation emulation from live deployment.

ThyroNet-XAI: An Explainable Multimodal Framework for Thyroid Cancer Analysis Using Clinical and Ultrasound Data

Thyroid cancer diagnosis requires accurate assessment of clinical characteristics and ultrasound imaging, yet existing artificial intelligence (AI) approaches often rely on a single data modality and provide limited interpretability. This paper presents ThyroNet-XAI, an explainable multimodal framework that integrates clinical tabular data and thyroid ultrasound images for automated thyroid cancer diagnosis and nodule analysis. The proposed framework employs SMOTE-ENN to address severe class imbalance in clinical data, followed by ANOVA-based feature selection to identify informative clinical attributes. A hybrid Random Forest–Deep Neural Network (RF-DNN) architecture is used for clinical classification, while an FCN–TransUNet architecture combines convolutional feature extraction with transformer-based global contextual learning for ultrasound nodule segmentation. To improve transparency, SHAP and LIME are incorporated for clinical predictions, while Grad-CAM provides visual explanations for image-based decisions. The framework is evaluated using the Kaggle thyroid clinical dataset, TN3K ultrasound dataset, and DDTI dataset for external validation. Experimental results reported in the study demonstrate 98% classification accuracy for the clinical branch, with the proposed hybrid clinical model achieving an F1-score of 0.7179 and ROC-AUC of 0.9803. For ultrasound analysis, the proposed segmentation framework achieves a Dice coefficient of 0.8974 and IoU of 0.8138, while multimodal imaging evaluation reports an AUC of 0.9942. These findings demonstrate the potential of combining multimodal learning, hybrid architectures, imbalance-aware preprocessing, and explainable AI to develop accurate and clinically interpretable thyroid cancer decision-support systems.

ECHO-LOCAL: Edge Context Orchestration for Privacy-Preserving Smart-Home Automation

ECHO-LOCAL introduces a privacy-preserving edge architecture that moves the entire smart-home presence estimation and actuation control loop onto a commodity OpenWrt router. By compressing a POMDP presence-control problem into a 5-state FSM estimator paired with a 100-byte 5×5 tabular Q-learning policy, it eliminates cloud round-trip dependencies and metadata leakage. The system achieves a fast 2.2 ms control loop and a compact 697 KB resident set size on live router hardware while enforcing zero WAN telemetry at the kernel level using nftables drop rules.

A Low-Cost ESP32 TinyML System for On-Device Arrhythmia Screening with Train-Deploy Preprocessing Parity

This paper presents an ultra-low-cost, privacy-preserving edge AI system for continuous arrhythmia screening deployed on an ESP32. The primary contribution is achieving train-deploy preprocessing parity by recreating the exact MIT-BIH dataset format on-device using a custom firmware pipeline and Pan-Tompkins R-peak detection. Additionally, the system features a hardware-specific filter chain optimized for local 50 Hz power grid interference and performs full on-device classification, ensuring that only scalar metrics—and no raw ECG waveforms—are transmitted to the cloud.

Spatial Image Steganalysis Using Fixed SRM Preprocessing and SRNet

1. A fixed 30-filter SRM preprocessing stage integrated into an SRNet-style backbone, replacing the learned front-end with handcrafted spatial residual filters.

2. Both configurations (BOSSBase alone and a combined BOSSBase+BOWS2 corpus ) are evaluated against SRNet and ResNet baselines across accuracy, AUC, precision, recall, and F1-score.

3. A paired augmentation pipeline that applies identical spatial transformations to corresponding cover-stego images, preserving their structural correspondence during
training.

4. An ablation study using an SRM+SRNet configuration to quantify the contribution of fixed SRM preprocessing relative to the baseline SRNet architecture.

5. Grad-CAM visualizations identify which image regions the model weighted most heavily when making classification decisions.

Toward Trustworthy Stroke Risk Prediction: A Balanced, Explainable and Externally Validated Machine Learning Framework

This study presents an interpretable and trustworthy stroke risk prediction framework that integrates Boruta–mRMR feature selection, training-only Borderline-SMOTE, comparative evaluation of nine classifiers, and an optimized CatBoost model with an OOF F2-based threshold. The framework further combines SHAP, LIME, and counterfactual explanations with fairness, calibration, robustness, bootstrap-based reliability, and cross-dataset external validation.

A Multi-Strategy Machine Learning Framework for Real-Estate Price Prediction: Pooled, Contextual, and Segmented Approaches

The primary contributions of this research are outlined below:
Empirical Comparison: An empirical comparison of three modeling strategies—a pooled baseline (Model A), a pooled contextual model (Model B), and dedicated per-segment models (Model C)—on 30,472 Bangladeshi property listings spanning four market segments, using five regression algorithms under a shared preprocessing and five-fold cross-validation protocol.
Statistical Testing: Paired, fold-level statistical testing with Holm–Bonferroni correction, confirming that the observed strategy differences are statistically supported.
SHAP-Based Interpretation: A SHAP-based model interpretation showing that the added market-context features are informative and that price determinants differ across the four segments.
Strategy Preference Evidence: Evidence on when each strategy is preferable: a pooled contextual model is strongest overall and on the sale segments, whereas dedicated segment models perform best on the rental segments.

EAM-XAI: Auditing Malware-Identifier Overlap with a Hybrid Behavioral–Retrieval Framework

Repeated memory snapshots carrying the same
dataset-provided malware identifier can cross record-level splits.
A high score may then measure recurring-identifier triage
rather than behavioral generalization to an unseen identifier.
We audit this dependence in CIC-MalMem-2022 and present
Evidence-Aware Malware XAI (EAM-XAI), a routed behavioral–
retrieval system: unseen or unavailable identifiers use a behavioral
LightGBM model, known pure identifiers use label-count retrieval,
and mixed-label identifiers use a local Extra-Trees specialist. The
filename-derived token is not treated as a verified executable hash.
Snapshot-stratified five-fold evaluation reaches 99.85% ± 0.06%
accuracy, whereas malware-identifier-disjoint evaluation reaches
85.49% ± 1.06%. A retrospective multi-snapshot extension that
requires joint access to an unseen identifier’s available captures improves a reproduced single-snapshot baseline from 85.50%±0.98%
to 89.93% ± 1.30%.

A Source-Grounded Multimodal Edge-Agent Framework for Inclusive Digital Service Access in Bangladesh

This research proposes a source-grounded multimodal edge-agent framework for improving access to digital services in Bangladesh. Its main contribution is the integration of Bangla/Banglish interaction, multimodal input, local edge inference, domain-specific routing, source-grounded retrieval, and risk-aware response selection within a single architecture. The work also demonstrates the feasibility of local execution through 120 software trials, achieving 92.5% overall success and 95.0% offline-continuity performance, while clearly distinguishing tested components from proposed future extensions.

Real-Time Detection of Toxic and Hate Speech in Esports Live Chat: A Comparative Study of Machine Learning and Deep Learning Approaches

The significant research contribution is the creation of the first multi-game annotated esports live-chat dataset for toxicity detection, containing 4,957 consensus-labeled messages from five major games. The study also provides a systematic comparison of eight ML, DL, and Transformer models, showing that DistilBERT achieved the best F1-score (0.8151), while Logistic Regression remained a competitive, much faster option for real-time moderation.

Beyond Balanced Benchmarks: A Deployment-Aware Evaluation of Prompt-Injection Detectors under Source Shift

This study evaluates three prompt-injection detectors under a frozen-threshold, deployment-aware setting.
It measures robustness across source shift, hard-benign prompts, and successful JailbreakBench artifacts without threshold retuning.
The results show substantial variation in recall, false-positive behavior, and cross-threat transfer across detectors.
The study also highlights deployment trade-offs through prevalence-aware precision, latency, and decision-margin analysis.

LEGCID: Lightweight Edge-Generated Classifier for Intrusion Detection on OpenWrt

Proposes LEGCID, an edge-native intrusion detection architecture deployed directly on commodity OpenWrt home gateways. The system trains a sparsified 30-feature random forest on IoT botnet traffic, compiles it into dependency-free C99 source via emlearn, and executes in-process flow classification directly from conntrack data with kernel-level nftables quarantine. It operates within strict embedded budgets—achieving 2.9 µs per-inference latency and a 772 KB resident set size while eliminating cloud telemetry egress and recurring compute costs.

A Non-Destructive IoT-Based Multi-Sensor System for Fruit Preservative and Surface Coating Detection

• Designing and developing a portable fruit sensing chamber
for controlled data acquisition.
• Design and integration of a multi-sensor platform, using
optical and gas sensing technologies.
• Co-ordinated data acquisition software and real-time
visualization.
• Production of a multi-modal dataset with synchronized
data from all sensors in a structured form.
• Development of a dataset capable of being used for future
fruit quality assessment systems using AI.

An End-to-End Deep Learning Framework for Animal Classification with Ensemble CNNs and Web Deployment

Animal-image classification is often evaluated as an
isolated model task, while the software path required to turn
a trained model into a usable system receives less attention.
This paper presents an applied deep-learning and softwareintegration workflow for closed-set classification of 30 animal
categories from a custom project corpus of 3,000 images.
The workflow combines image standardization, duplicate
screening before partitioning, a nominal class-wise 70/15/15
split, training-only augmentation, transfer learning, fixed classindex alignment, heterogeneous CNN fusion, and a browseraccessible Flask inference service. The deployed ensemble combines MobileNetV3Small, EfficientNetV2B0, and DenseNet121
through equal-weight averaging of aligned softmax outputs.
On a 448-image development validation set, the observed
development result was 98.44% accuracy with macro precision,
recall, and F1 of 99%, 98%, and 98%, respectively.

ScopeGuard: Mitigating Goal Drift and Confused Deputy Attacks in MCP Agent Pipelines

This paper presents ScopeGuard, a middleware-based security architecture for Model Context Protocol (MCP) agent pipelines that jointly addresses goal drift and confused-deputy attacks. ScopeGuard mediates planner-generated tasks before execution using four complementary checks: semantic intent alignment, provenance verification, least-privilege permission validation, and prompt-injection detection. The prototype is evaluated on 831 labeled tasks, including real goal-drift trajectories and benchmark injection and confused-deputy attacks, demonstrating 80.6% detection of real goal drift and 100% detection of the evaluated injection and confused-deputy cases.

MindBridge: A Multi-Modal AI Platform for Depression and Anxiety Risk Stratification

1. Developed an integrated multi-modal AI platform combining self-report instruments, ClinicalBERT free-text triage, facial affect classification, and EHR vitals with an XGBoost risk ensemble and SHAP attribution.
2. Implemented a 30-day Temporal Fusion Transformer (TFT) to accurately forecast PHQ-9 trajectories and symptom progression.
3. Designed a retrieval-augmented CBT chatbot featuring explicit safety refusal triggers and crisis-line escalation pathways (0 harmful responses on a 200-prompt benchmark).
4. Validated on a retrospective cohort of 4,820 patients, achieving state-of-the-art performance (AUROC 0.892, AUPRC 0.741, F1 0.842) with comprehensive demographic bias analysis and HIPAA-compliant architecture.

X-CRSAA: An Explainable, Clinically-Regularized Self-Attention Autoencoder for Unsupervised Thyroid Anomaly Detection

The core contributions of this work are as follows:
1) A Feature-Tokenized Multi-Head Self-Attention (FTMHA) encoder paired with dual-decoder reconstruction
heads, optimized via a composite loss enforcing biological TSH/FTI constraints (Sections IV and V).
2) A Peaks-Over-Threshold EVT threshold calibration
pipeline that replaces heuristic percentile cut-offs with
extreme-quantile estimation (Section VI).
3) A dependency-free Kernel-SHAP explainability module
providing global and local clinical feature attributions
validated against synthetic ground truth (Section VIII).
4) An extensive empirical ablation demonstrating that
reconstruction-only attention optimization outperforms
masked-feature pretraining and latent-density fusion on
tabular medical data (Section XII).

Sequential Modeling of Corporate Bankruptcy Risk Using Hidden Markov Model Augmented Transformers.

This study contributes a hybrid framework for bankruptcy risk prediction that integrates Hidden Markov Models to capture latent temporal financial dynamics with Transformer-based models for multivariate financial time-series classification. The study also provides a systematic comparison with classical machine learning baselines, demonstrating the potential of temporal representations for identifying bankruptcy risk under severe class imbalance.

Rice Yield Prediction and Crop Health Assessment in Bangladesh Using Satellite Imagery and Machine Learning

This study proposes an integrated machine learning framework for rice yield prediction and crop health assessment in Bangladesh using satellite imagery and environmental data. The study combines NDVI derived from Sentinel-2 satellite imagery with rainfall and temperature data to characterize rice-growing conditions. Machine learning models are employed to predict rice yield and assess crop health across different rice seasons in Mymensingh and Khulna districts. The proposed approach demonstrates the potential of integrating remote sensing and machine learning for data-driven rice monitoring and yield estimation in Bangladesh.

Explainable Calibrated Hierarchical XLM-R for Bilingual PHQ-9 Depression Severity Classification

The study proposes a calibration-aware hierarchical XLM-R framework for five-level PHQ-9-guided depression severity classification in Bangla and English. Its key contributions are hierarchical modeling of three open-ended responses with ordinal-aware learning, bilingual performance analysis, improved confidence reliability through temperature scaling, and interpretable prediction analysis using hierarchical attention and LIME.

Latent-Space Counterfactual Explanations for Cervical Cytology Classification

Deep convolutional networks now classify cervical
(Pap smear) cytology with high accuracy, yet clinical adoption
is held back not by accuracy but by opacity: saliency methods
such as Grad-CAM and SHAP show where a network looked,
but never what would need to change for the decision to differ.
Treating accuracy as a solved sub-problem, this paper contributes
a counterfactual explanation module for Bethesda-system cervical
cytology classification, presented as a pilot feasibility study. We
first select a backbone in a data-driven manner from four
CNNs (ResNet-50, EfficientNet-B3, DenseNet-121, MobileNetV2)
and fuse the top two, DenseNet-121 and EfficientNet-B3, via a
softmax-gated adaptive cross-scale fusion module with a Convolutional Block Attention Module (CBAM), reaching 95.45%±1.37%
test accuracy and 0.9506±0.0135 macro-F1. This classifier is then
frozen and used purely as a differentiable oracle for our main
contribution: a convolutional variational autoencoder coupled
with a latent-space optimization procedure that searches for a
constrained, class-flipping, background-preserving perturbation
of a cytology image toward a target class. Across four diagnostically important Bethesda transitions (n = 20), the module
achieves a 100% validity rate, mean SSIM of 0.549, mean
LPIPS of 0.625, mean sparsity of 0.403, and mean target-class
confidence of 0.985, while keeping background disturbance low
(|D| = 0.082), indicating that most of the edit is confined to
the cell-bearing foreground. The resulting difference maps are
qualitatively consistent with, yet informationally complementary
to, Grad-CAM++ and SHAP saliency.

An Edge-AI Driven Multi-Layered Air Defense Framework for Real-Time Cost Asymmetric Threat Prioritization

The primary contributions of this paper include developing a novel Cost-Asymmetric Threat Prioritization (CATP) algorithm that merges kinematics, trajectory alignment, and CNN-derived visual features (payload severity and destructibility) to resolve multi-target tie-breaking. Additionally, the framework introduces a Cost-Asymmetric Ratio (CAR) that achieves a 32.0% reduction in interceptor costs via a tri-layered response strategy, all while being fully validated on resource-constrained edge hardware achieving an end-to-end latency of 145.33 ms using only 3.5% CPU and 14.3% RAM.

Wind Turbine Blade Structural Health Monitoring Using Piezoelectric Sensors and Machine Learning

This paper presents a machine learning framework
for classifying fatigue-induced damage stages in a wind turbine
blade using piezoelectric (PZT) guided-wave measurements. A
cycle-aware, group-based data-splitting strategy is employed to
prevent measurements from the same fatigue cycle from
appearing in both training and test partitions. Four classifiers –
Random Forest, XGBoost, Logistic Regression, and Support
Vector Machine are trained using 44 time- and frequencydomain features extracted from four PZT receiver channels at
six excitation frequencies (20–120 kHz). The dataset comprises
10,968 feature vectors spanning 1,828 unique fatigue-cycle
instances from a Sonkyo Windspot 3.5 kW composite blade
subjected to cyclic loading. Four structural states are
considered: Pre-crack, Crack-3mm, Crack-10mm, and Postoverload. On the held-out test set of 2,196 feature vectors,
Random Forest and XGBoost achieve accuracies of 99.64% and
99.68%, respectively, with AUROC values of 0.9996 for both
models. Logistic Regression and Support Vector Machine
achieve accuracies of 98.00% and 96.68%, respectively.
TreeSHAP analysis identifies dispersion- and energy-related
features from two PZT channels as consistently influential
across the tree-based models. A label-permutation control
reduces accuracy to 34.6% when training labels are shuffled,
supporting the conclusion that the observed classification
performance depends on the original damage-label structure
rather than the partitioning procedure alone. The results
demonstrate the feasibility of machine-learning-based fatiguestage classification on a single benchmark specimen, while crossspecimen generalization remains an open question.

Efficient Wound Segmentation Using Lightweight U-Net with Adaptive Skip Fusion and Wound-Specific Attention Mechanisms

The main contributions of this work are summarized
as follows:
• Adaptive Skip Fusion (ASF), Wound-Specific Attention
(WSA), and Progressive Refinement (PR) are all inte-
grated into a single U-Net architecture in this lightweight
wound segmentation system.
• In order to retain computing efficiency while improving
the depiction of heterogeneous wound tissues and uneven
boundaries, a wound-oriented attention mechanism is
presented.
• With a Dice coefficient of 0.7584 and an improvement
in segmentation performance of 4.61% over the baseline
lightweight U-Net, the efficacy of the suggested frame-
work is demonstrated on a public wound segmentation
dataset.

An Educational RTL-to-GDSII Flow for a Non-Pipelined 8-bit Accumulator-Based CPU Using Open-Source Tools

Designed and implemented a complete open-source RTL-to-GDSII flow for an 8-bit accumulator-based CPU, verifying six sub-blocks individually before top-level integration, achieving zero DRC/LVS errors and clean timing closure using Yosys, OpenLane 2, and the SKY130 PDK.

Paper Type: Full Research Paper