Paper Submission - NPS Australia Submission System

LightSkinNet: Protocol-Aware Skin Lesion Classification with Segmentation-Guided Networks

Automated dermoscopic analysis is commonly approached as two separate tasks: lesion segmentation and skin lesion classification. However, the wide range of reported accuracies on the HAM10000 dataset is difficult to interpret because different studies often use different evaluation protocols. In this work, we introduce LightSkinNet, a two-stage framework in which a lightweight segmentation model first localizes the lesion, and the resulting cropped image is then passed to a classifier. We investigate two encoder designs consistently across segmentation and classification: a convolutional encoder incorporating squeeze-and-excitation and bottleneck attention mechanisms, and a hybrid architecture with a transformer-based bottleneck. On ISIC 2018 Task 1, the convolutional model achieves a Dice coefficient of 0.8927, while on HAM10000, our best configuration achieves 0.9271 accuracy under an image-level split, with the convolutional design again outperforming the hybrid model (p < 10^-12, McNemar's test). We further examine the evaluation protocol and find that the 10,015 HAM10000 images correspond to 7,470 distinct lesions, with 661 lesions crossing the training-test boundary under the commonly used image-level split. We also show that melanoma detection is strongly influenced by the decision threshold rather than solely by the learned representation: validation-only threshold tuning reduces missed melanomas from 49 to 22 without modifying the model parameters. To assess whether the classifier focuses on clinically relevant regions, we quantify the overlap between Grad-CAM activations and the predicted lesion masks. Finally, we report both the strengths and limitations of the proposed framework, including the finding that a low parameter count does not necessarily translate into low computational cost.

Energy Sustainability in Bangladesh: Evaluation of Global Models and Future Directions

Globally, sustainable energy management has got utmost priority irrespective of economies of scale. It is the melting pot where optimum balance between economy, energy and environment can be ensured. On the contrary, Bangladesh is a disaster-prone country, confronting with multiple challenges of economic emancipation and its impact on escalated energy growth, and subsequent environmental concerns, that ultimately creating another big challenge for energy sustainability. Like many other countries, Bangladesh is not an energy independent country, but sets vision to serve the nation with reliable and affordable electricity. At this crossroads, sustainable energy management is not an option, but hard reality for Bangladesh. To unfold the sustainability issues, the study adopted pragmatic worldview as research approach, where both exploratory and explanatory methods were applied. The study compared Bangladesh’s energy model with the energy models of USA, EU, Nordic, China, India, Vietnam, Indonesia and Pakistan. The study identified supply-demand imbalance, absence of least cost generation mixes with renewable and clean energy integration, lack of optimum utilization of available resources through appropriate technologies and skilled manpower, and insufficient cash flow to run the sector efficiently, in Bangladesh’s energy model, thereby exposed the energy model to vulnerabilities. Based on the outcomes, the study recommends a comprehensive and sustainable energy model, where national energy policy will be inclusive, data-driven and stakeholders-oriented.

Robustness and Privacy Trade-offs in Federated Multimodal Brain Tumor Segmentation: Missing Modalities, Differential Privacy, and Membership Inference

A companion paper builds a federated learning
pipeline for brain tumor segmentation. It trains across four
simulated hospitals with uneven data, and one hospital is always
missing a scan type. It also adds security: TLS, node checks, and
secure aggregation. That paper leaves three questions open. Does
a proximal FedAvg variant (FedProx) help with the uneven data?
Does a model built for missing scans really recover the accuracy
a normal model loses? What does adding formal differential
privacy (DP) cost, both in accuracy and in real protection? This
paper answers all three, and adds a fourth question: does DP
actually stop a real privacy attack, or does it just cost accuracy
for nothing? We find that FedProx performs worse than plain
FedAvg at both dataset sizes we test. We build a modality-aware
fusion model, trained to handle missing scans. It matches the
normal model’s accuracy when every scan is present. But when
the FLAIR scan is missing – the most important scan in our
data – it recovers 0.26 Dice points (0.4785 → 0.7382) that the
normal model loses. On privacy, we find a real problem. Our
DP method adds noise to a ∼971K-parameter model update.
The amount of noise needed scales with the square root of the
parameter count, not with the clipping norm. Every noise level
that keeps accuracy usable gives a privacy guarantee (ε) so large
it means almost nothing. Every noise level with a meaningful
ε (around 16–111) breaks the model completely. We also run a
membership-inference attack, which checks whether an attacker
can tell if a patient’s scan was used in training. This attack
barely works at all (AUC 0.50–0.52), even with no DP applied.
So the accuracy DP costs is not clearly buying any real protection
against this attack. We report all four results honestly, including
the two that did not work as hoped, because a federated system
that is honest about where privacy tools fail is more useful than
one that only reports wins.

Computer Vision-based Soil Health Analysis using Machine Learning

1. We design an attention-based cross-modal fusion strategy
that enables adaptive interaction between heterogeneous
data sources, improving the learning of complementary
spectral, spatial, and environmental representations.
2. We introduce a normalized and interpretable Soil Health Index (SHI) derived from dual-standard pH predictions, translating continuous outputs into actionable categorical soil
health classes for practical agricultural decision-making.
3. We incorporate computationally efficient training strategies,
including mixed-precision learning, stochastic weight averaging, and data augmentation, to enhance robustness while
maintaining scalability for large-area soil monitoring.
4. We conduct an extensive evaluation on the pan-European
LUCAS 2018 soil dataset, demonstrating strong generalization across diverse soil types, climatic regions, and land-use
categories.

Economic Prospects and Challenges of Solar Energy Adoption in Bangladeshi Households

This study provides recent household-level evidence on solar energy adoption in Bangladesh using primary data from 166 households. It finds a 41.6% adoption rate and identifies knowledge of solar technology as the only significant predictor of adoption (OR = 1.51, p = 0.016). The study also highlights key barriers, particularly high installation costs, weather-related efficiency concerns, and maintenance problems. The findings suggest that accelerating household solar adoption requires not only financial support but also awareness programmes, technical training, and reliable after-sales services. These results provide timely and practical policy insights for expanding decentralized renewable energy in Bangladesh.

An Efficient Approach to Beetle Damage Detection of Banana Using Deep Learning

This study develops a lightweight YOLOv8n-based framework for instance-level detection of visible banana surface damage using a locally collected and manually annotated dataset from Bangladesh. The model distinguishes healthy and unhealthy banana instances and achieves 87.04% mAP@0.50 on the held-out test set while maintaining a compact 5.96 MB model size and fast inference, making it suitable as a practical baseline for automated banana surface-quality screening.

Rice Leaf Disease Classification Using Fine-Tuned ResNet152V2: Benchmarking CNN Models

This study presents a fine-tuned ResNet152V2-based transfer learning model for six-class rice leaf disease classification using a balanced dataset of 2,100 images. The proposed approach incorporates standardized preprocessing, stratified data splitting, and controlled training-time augmentation to improve generalization and reduce overfitting. A systematic comparison with VGG16 and MobileNetV2 demonstrates the accuracy–efficiency trade-off among the evaluated architectures. The proposed FT-ResNet152V2 achieves 96.24% test accuracy, with 0.96 precision, recall, and F1-score, outperforming VGG16 and MobileNetV2. Comprehensive evaluation using confusion matrices, ROC curves, and precision–recall curves further demonstrates its strong classification capability, while model complexity analysis highlights the computational cost associated with its higher accuracy.

RLCB-Net: A Lightweight Attention-Based Network for Rice Leaf Disease Detection and Classification

This study proposes Hybrid RLCB-Net, a lightweight and computationally efficient deep learning model for rice leaf disease detection and classification. The model is evaluated using single-, dual-, and multi-source datasets and compared with established architectures, including ResNet50, EfficientNet-B0, MobileNetV2, SqueezeNet, and VGG16. A comprehensive evaluation is conducted using accuracy, precision, recall, F1-score, inference latency, throughput, model size, parameters, and GFLOPs. In addition, SHAP-based explainable AI is employed to interpret model predictions and identify influential disease-related regions. The proposed model demonstrates a strong balance between classification accuracy, computational efficiency, and inference speed, making it suitable for real-time rice disease detection in resource-constrained agricultural applications.

Adaptive Cardiovascular Disease Screening via Uncertainty-Guided Clinical Feature Acquisition

Cardiovascular disease (CVD) remains a major global health challenge, underscoring the need for reliable and efficient artificial intelligence-based screening strategies. Conventional CVD prediction models generally employ a fixed set of patient attributes and produce definitive classifications without explicitly considering predictive uncertainty or the resource burden associated with acquiring additional clinical information. Building on an empirical analysis of a public dataset containing 70,000 patient records, this study proposes a conceptual framework for uncertainty-guided adaptive CVD screening. The empirical foundation includes clinically informed feature engineering using body mass index (BMI), mean arterial pressure (MAP), and blood-pressure stage, together with SelectKBest Chi-Square feature selection, principal component analysis (PCA), and the evaluation of seventeen machine learning, deep learning, and CNN-based hybrid models. Multilayer Perceptron (MLP) and CNN+XGBoost achieved the highest benchmark performance, each attaining 96\% accuracy and a 0.96 F1-score. The proposed framework extends this fixed-feature paradigm by organizing clinical variables into resource-aware acquisition stages and estimating uncertainty after each stage. Additional information is acquired only when necessary, while persistently uncertain cases may trigger abstention and clinical referral. Reliability is further conceptualized through discrimination, calibration, uncertainty quality, robustness, and subgroup consistency, with future empirical validation required. This design prioritizes patient-specific information needs while minimizing unnecessary clinical measurements overall.

Vision-Language Representation Learning for Image Captioning with Attention-Based Interpretability

Recent advances in vision–language models have significantly improved image captioning by enabling more effective learning of visual and textual representations. Despite these improvements, understanding how transformer-based models generate captions remains challenging because their decision-making process is often difficult to interpret. This paper presents a vision–language representation learning framework for image captioning that integrates a Vision Transformer (ViT) encoder with a BERT-based decoder to generate descriptive image captions while incorporating attention-based interpretability. The proposed framework employs attention visualization and head-wise attention analysis to identify the image regions that contribute to caption generation, providing greater insight into the model’s internal reasoning process. Experimental evaluation on the COCO Image Caption dataset achieved BLEU-1, BLEU-2, BLEU-3, and BLEU-4 scores of 0.6384, 0.4636, 0.3383, and 0.2516, respectively. In addition, a Gradio-based interactive web application was developed to enable real-time caption generation together with attention map visualization. The experimental results demonstrate that the proposed framework generates semantically meaningful captions while improving the transparency and interpretability of transformer-based image captioning systems.

ScanRig-Edge: Leakage-Aware Browser-Native Temporal Pose Recognition for Real-Time Home Exercise Coaching

Real-time home exercise coaching requires recognition that is accurate under pose variation while remaining
efficient enough for client-side execution without continuous
cloud video processing. This paper presents ScanRig-Edge, a
browser-native home-exercise coaching application that integrates webcam-based MediaPipe pose tracking, temporal exercise
recognition, confidence-aware unknown rejection, and real-time
user feedback. Its recognition engine converts pose landmarks
into 48-frame normalized sequences and classifies nine exercises
plus an unknown class using a compact one-dimensional con
volutional network. A 2,200-sequence synthetic benchmark is
constructed from 80 persistent profiles with five profile-disjoint
split manifests and an integrity audit. The audit finds zero exact
duplicate sequences and zero train/validation/test profile overlap.
Under a five-seed matched-augmentation protocol, ScanRig-Edge
achieves a mean Macro-F1 of 99.53% with a standard deviation
of 0.33%, and a mean unknown-class F1 of 99.34% with a
standard deviation of 0.60% on post-freeze synthetic distribution
shifts. Hformer-A reaches the highest mean Macro-F1 of 99.92%,
but ScanRig-Edge achieves 0.595 ms median and 0.880 ms p95
browser inference under ONNX Runtime Web/WASM. These
measurements place ScanRig-Edge on the observed accuracy
latency Pareto frontier. The classifier is integrated into a working
React/Vite application with local inference, confidence-based
unknown rejection, model hot-reloading, and a privacy-conscious
pose-sequence collection workflow for future participant-disjoint
real-user validation.

JEPA-TTA -Test-Time Adaptation via Latent Prediction for Domain-Shifted Medical Images

Contributions
The main contributions of this study are as follows:
• We propose JEPA-TTA, a single-image test-time adap-
tation framework that introduces latent-space predictive
learning as an adaptation objective for cross-domain
medical image classification.
• We develop a masked-view adaptation mechanism that
combines latent predictive consistency with entropy reg-
ularization, while restricting parameter updates to the
predictor and LayerNorm affine parameters.
• We systematically evaluate the proposed framework
against Source-Only, AdaBN, TENT, SHOT, and MEMO
under a CheXpert-to-COVID-19 Radiography Database
domain shift.
• We conduct component-level ablation experiments to
investigate the contributions of JEPA pretraining, en-
tropy regularization, restricted parameter updating, and
the number of target views.
• We analyze both discriminative performance and feature-
space representations to characterize the behavior of
JEPA-TTA under the evaluated domain shift.

Leakage-Controlled Evaluation of Tree-Based Models for Binary Intrusion Detection

This paper shows that near-perfect aggregate IDS metrics can mask complete failure on a supported attack category. Using 60 reproducible, leakage-controlled experiments, it establishes transparent partitioning and per-category reporting as essential for valid ML-based IDS benchmarking.

From Hierarchy to Participation: How Leadership Shapes Budgetary Information Systems

Abstract—The purpose of this research is to understand the relationship between hierarchical level and the leadership style of an organization’s executives and its budgetary system. This study will also investigate the influence of other behavioral factors (information asymmetry, procedural fairness) and employees’ organizational commitment. Managers at different levels are involved in the budgetary system, each with different responsibilities. Their leadership style in helping subordinates achieve targets also influences the process. The present study aims to understand this scenario in the context of manufacturing companies in Bangladesh. For this study, the term budgetary system’ is conceptualized in terms of subordinates’ participation. This study excludes the technical aspects of budgeting, as the current literature has extensively studied those variables. Data will be collected through a Likert-type questionnaire survey. The study will measure the constructs by adapting scales from earlier studies. A Structural Equation Model will be used to measure the relationships among the variables. The expected outcomes of the study are: executives with higher responsibility and a supportive leadership style encourage subordinates’ participation in the budgetary process; second, the relationship between management level and leadership style of executives with participation in the budgetary system is moderated by the extent of information asymmetry and procedural fairness of budgeting; and third, employees who participate in the budgetary sytem feel increased level of organizational commitment.

Index Terms—budget, participation, management level and leadership style of executives, manufacturing companies

A leak free re-evaluation of augmented Alzheimers mri with deep neural network models

Solar Power Prediction in Coastal Zones of Bangladesh: A Hybrid Empirical-Machine Learning Approach for the Teknaf 20 MW Solar Power Plant

This study reports a relative assessment of single machine learning, ensemble, and hybrid deep learning methods for multi-step solar radiation and power forecasting tailored to the coastal climate of Bangladesh. Astronomical solar geometry, empirical global solar radiation regression models (Angstrom-Prescott, Akinoglu and Ecevit, Ampratwum and Dorvlo, and Newland), and five core algorithms (XGBoost, Random Forest, ANN, LSTM, and SVR) are applied to the Teknaf 20 MW solar power plant in Cox’s Bazar, Bangladesh. The models predict daily and multi-step power generation using meteorological and engineered features obtained through feature selection. The empirical regression formulations successfully capture seasonal variations with high correlation coefficients (r>0.98). Among the appraised models, Random Forest Regression demonstrates the superior performance (R2 = 0.9910, RMSE = 0.1898 MW, MAPE = 4.70%, Accuracy = 95.30%), closely followed by XGBoost and SVR. These tree-ensemble and kernel-based methods provide accuracy comparable to or exceeding complex deep learning models at a low computational cost, making them highly suitable for solar power planning and grid management in tropical and monsoon regions.

Segmentation-Guided Attention Transfer Learning for Breast Cancer Classification in Digital Mammography

1. An annotation-free attention-weighting pipeline combining CLAHE and ROI extraction into a pixel-level input attention map;
2. A two-phase fine-tuning protocol across three architectures under a patient/exam-level, group-aware split, closing a leakage gap present in an earlier image-level evaluation; and
3. A quantitative validation of the generated attention mask against radiologist-drawn ROI annotations (Dice/IoU), complementing the qualitative Grad-CAM++ analysis.

Uncovering Healthcare Concerns Among Bangladeshi Patients Seeking Medical Care in India Using BERTopic

Social media has become an important source of health information and peer support for patients seeking medical treatment. This study aimed to identify and characterize the major information needs, healthcare concerns, and discussion themes of Bangladeshi patients seeking medical treatment in India. This study examines the main discussion themes within two Facebook communities, “CMC Hospital Vellore Treatment Community Bangladesh” and “Bangladesh Treatment Community for Indian Treatment”. Data were scraped from publicly available Facebook posts (n = 1500) using web scraping. BERTopic was applied to analyze the data. This pattern yielded 28 interpretable topics, with coherence scores from 0.2990 to 0.8905. The results also found conversations regarding cardiovascular diseases, kidney diseases, neurological disorders, cancer, infertility, spinal disorders, ophthalmic disorders, liver diseases, gynecological disorders, and gastrointestinal diseases. Additionally, several other non-medical themes were noted, viz., treatment expenses, accommodation, travel, and procurement of medicines, as well as tourism planning, indicating the practical difficulties encountered by patients in a foreign country when seeking healthcare in India. The results reveal that in addition to being a source of medical information, these Facebook groups act as support networks in which patients turn for helpful advice such as what to expect, share personal experiences, and seek emotional support. The findings may assist researchers, clinical teams, and policymakers to gain insights into the information needs of cross-border patients, helping to enhance the provision of patient-centred healthcare services.

FairMoE-Surv: A Fairness-Gated Neuro-Symbolic Mixture-of-Experts for Equitable Clinical Risk Prediction with Cross-Cohort Validation

Abstract—Fairness-aware clinical risk models are often evaluated using aggregate metrics that may obscure disparities across
protected groups and provide limited evidence of generalizability.
We address both issues with FairMoE-Surv, a neuro-symbolic,
fairness-gated mixture-of-experts (MoE) architecture that combines a hybrid learned/symbolic router, clinically grouped mechanism tokens, momentum-updated population-prototype attention,
a top-2-of-6 MoE layer, demographic-parity and proxy-feature
consistency regularizers, and a calibration module. On N =
28,800 real, de-identified CIBMTR allogeneic hematopoietic-celltransplantation records spanning six self-reported race groups,
FairMoE-Surv achieves 89.2% AUROC, 88.9% accuracy, 88.4%
F1, 89.1% precision, 87.8% recall, and 88.7% concordance
index, outperforming CoxPH (83.8% AUROC), Random Survival
Forest (85.2%), DeepSurv (86.1%), DeepHit (84.8%), and MTLR
(86.9%). It also attains the lowest demographic-parity gap (9.6%
vs. 14.1–20.1%), equal-opportunity gap (9.1% vs. 15.3–23.6%),
and expected calibration error (2.8% vs. 4.1–6.7%) across all
six race groups. A five-component leave-one-out ablation shows
that every component contributes: removing the MoE layer
reduces AUROC to 84.6%, while removing the calibration module
reduces it to 86.9% but increases ECE to 3.5%. Retraining the
identical architecture on a disease-mismatched SEER breastcancer cohort (N = 4,024, three race groups) preserves FairMoESurv’s superiority but narrows its fairness margin from 27–46%
to 7–9%. These findings support cross-cohort retraining as
essential for validating fairness generalization.

SpearXAI-Org: Target-Context-Aware Benchmark and Explainable Hybrid Reference Framework for LLM-Assisted Spear-Phishing Detection

Our paper’s biggest contribution isn’t a smarter phishing-detector. In fact, we openly admit that our hybrid classifier itself doesn’t beat a simpler one in any measurable way. What we actually contributed is:

1. A better test set. We built 8,200 emails split into 5 clear types — normal email, generic scams, human-written targeted scams, AI-written targeted scams, and AI-reworded targeted scams. Most past research only asked “scam or not,” lumping AI-written attacks in with everything else. This lets us (and others) test tools specifically against AI-written phishing.
2. A more honest way to test these tools. We checked for common mistakes other studies make — like accidentally letting the same email (or a reworded copy of it) appear in both the “training” and “testing” piles, which makes a tool look smarter than it really is. We fixed that.
3. A real-world reality check. We showed that even a tool that scores 99% in a clean lab test would still flag mostly innocent emails as “false alarms” once used in a real company inbox, where actual scams are rare. This is a really important, often-ignored point.
4. Proof that “fancier” doesn’t always mean “better.” We tested whether combining two AI methods, or adding extra explanation features, actually improved accuracy — and found it mostly didn’t. That’s a useful, honest finding, since many papers overstate how much these extra bells and whistles help.

Bottom line: Our real contribution is a fair, well-tested “rulebook and test set” for judging future phishing-detection tools — not a new, better detector itself.

CMSA-Net: Cross-Modal Sparse Attention and Generative Representation Repair for Multimodal Classification Under Missing Inputs

Abstract—Heterogeneous clinical data streams—electronic
health records (EHR), computed tomography (CT), magnetic
resonance imaging (MRI), and wearable Internet-of-Things (IoT)
signals—offer complementary information for cardiovascular
risk stratification but are often incomplete at inference time.
We propose CMSA-Net, a novel Cross-Modal Sparse Attention
Network to integrate data from multiple modalities using a
transformer encoder with a specific missing modality mask
and top-K sparse cross-modal attention mechanism. In order
to deal with missing modalities arbitrarily, CMSA-Net has a
conditional variational auto-encoder based Generative Modality
Repair (GMR) mechanism that generates robust embeddings
for missing modalities along with a Monte-Carlo Dropout head
for accurate uncertainty quantification. On the standardized
multimodal dataset of 738 encounters, CMSA-Net delivers an
outstanding performance of 0.979 AUROC (bootstrapped 95%
CI: [0.976, 0.982]), 0.972 AUPRC, 0.909 F1, and 0.923 accuracy,
beating the best baseline, an early-fusion MLP, by 0.031 AUROC
and 0.058 F1. Comprehensive robustness analysis confirms invariance of performance across random missing modalities with
AUROC change by less than 0.001 with up to 70% modality
dropout. Moreover, five-fold cross-validation results in average
AUROC 0.983 ± 0.004 with well-calibrated expected calibration
error 0.040. These results establish CMSA-Net as a highly
effective, resilient architecture for robust clinical prediction under
missing-data conditions.

Deep Learning Based Breast Cancer Detection With Web Application

Breast cancer is one of the common cancers for
women in the world. It is also a leading cause of cancer-related
deaths.Early and accurate detection can improving the chances
of recovery for patients.In hand ,it also depends a lot of the
experience for the pathologist.Recent progress in AI and deep
learning has made computer diagnosis more effective.Neural
Networks or CNNs can automatically find patterns in complex
medical images.Deep learning methods have also shown promise
in classifying.ResNet-based models can learn features better
thanks to residual connections.In this study introducing a deep
learning system designed to detect breast cancer from images.The
system classifies images into two categories Malignant and
Benign.Our framework combines MobileNetV3 and ResNet50V2
to capture types of image features that work well together.Using
crossvalidation and Test Time Augmentation TTA to increase the
reliability of predictions.Apply Explainable Artificial Intelligence
or XAI techniques to highlight the parts of the image for the
decision.At the end building a web application to allow real-time
predictions using the trained model.

Leakage-Controlled Tomato Leaf Disease Classification Using Attention-Guided CNN and GCN Feature Fusion

Accurate tomato leaf disease recognition can be
overstated when visually repeated images occur across data
partitions or evaluation is restricted to familiar sources. This
study presents a leakage-controlled classification framework that
combines EfficientNetV2-S, the Convolutional Block Attention
Module (CBAM), an inductive Graph Convolutional Network
(GCN), and feature-level fusion. Starting from 31,450 cleaned
images representing 11 tomato leaf classes, a perceptual-hash
audit identified 1,237 repeated-image groups involving 2,539
images. After removing 24 cross-class conflicts and one corrupted
image, 31,425 images were partitioned group-wise into 21,990
training, 4,718 validation, and 4,717 held-out test samples, with
no verified SHA-256, perceptual-hash, or source-group overlap
across the partitions. The attention-enhanced network produced
1,280-dimensional image embeddings, while a cosine-neighbor
graph with k = 5 and an inductive graph network generated 128-
dimensional relational features.

Embedded Fail-Safe and Real-Time Communication Framework for HTTP-Based Wireless Mobile Robot Teleoperation

The research proposes an Adaptive Communication-Aware Autonomous Motion Termination (Adaptive-CAAMT) topology for HTTP-based wireless mobile robot teleoperation. Its key contribution is the integration of adaptive timeout supervision, network-quality-based motion degradation, heartbeat synchronization, timestamp validation, and autonomous fail-safe motor shutdown to improve communication resilience and robotic safety. Experimental results demonstrate 100% fail-safe shutdown success across tested fault conditions, with emergency stopping achieved in approximately 143–145 ms.

An Intelligent Software Project Risk Prediction Engine Using Machine Learning and Web-Based Decision Support Systems

The significant contribution of this research is bridging the gap between theoretical machine learning models and practical industrial application in software project management. It develops an end-to-end Intelligent Software Project Risk Prediction Engine using a Random Forest classifier and seamlessly integrates it into a real-time, Django-based Web Decision Support System (DSS). This integration provides project managers with an accessible, user-friendly platform to dynamically assess early-stage project risks based on 51 historical project features, overcoming the limitations of traditional manual assessments and standalone “black box” ML scripts.

MetaTrojan: Backdoor Attacks on Metadata-Conditioned Medical Diffusion Reconstructors

MetaTrojan is the first backdoor attack on a metadata conditioned medical diffusion reconstructor that operates purely through the text conditioning channel, requiring no pixel domain access, and is therefore structurally invisible to all existing image domain defenses. It demonstrates that a single poisoned metadata token, injected by one malicious actor during data curation, is sufficient to reliably hallucinate clinically plausible lesions at inference time, while remaining statistically indistinguishable from a benign model on clean inputs. A lightweight whitelist sanitization layer is proposed as a defense, reducing ASR from 100% to approximately 0% with no measurable cost to reconstruction fidelity (30.77 dB vs. 30.74 dB benign baseline).

Architecture of a Resilient Root-Level IoT Communication Infrastructure for Multi-Hazard Natural Disasters in Bangladesh

This research presents an affordable, resilient multi-hazard early warning system engineered for flood-prone, low-resource regions like Bangladesh. Running on an ATmega2560 microcontroller, the system integrates water clearance, liquid submersion, combustible gas, and ambient light sensors to activate local hazard alarms in under 15 ms. To prevent communication blackouts when physical broadband connections fail, it features an autonomous dual-gateway engine that switches from wired Ethernet to cellular GSM in just 3.42 seconds. Crucially, the platform bridges rural populations using basic 2G feature phones through an edge-level parser that extracts location codes and trapped headcounts from simple SMS tokens (e.g., UNSAFE#FENI01#08) with 99.4% accuracy, without needing smartphones or internet access. These aggregated reports feed an operations dashboard that automatically dispatches targeted SMS alerts to rescue teams, healthcare units, and local administrators. With a total hardware cost under $45.00 USD, this design offers an inclusive, dependable, and life-saving disaster telematics solution.

A Deep Learning Framework for Cybercrime Topic Discovery and Sentiment Analysis in Bangladeshi Facebook Posts

This integrated approach provides a clearer understanding of cybercrime-related discussions and public perceptions in Bangladesh.

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 mon itoring and controlling prototype that was designed and de veloped 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 water level 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.

SeedNet: A Lightweight Fine-Tuned Deep Learning Framework for Plant Seedling Classification in Precision Agriculture

Developed a lightweight deep learning framework, EfficientNetB0, based on EfficientNet model, with transfer learning and selective fine-tuning to classify visually similar plant seedlings with accuracy. The proposed two-stage training technique is effective for domain adaptation and generalization, and achieves 95.40% test accuracy, 94.54% Macro-F1, and 99.90% micro-ROC-AUC on the V2 Plant Seedlings Dataset, without compromising the computation efficiency. The work highlights the possibility of practical and automated weed identification for precision sustainable agriculture.

Deep Learning-Based Gleason Grading of Prostate Cancer detection from Histopathology Images

The Gleason grading of prostate cancer is well
known for being subjective in clinical practice, interpathologist
disagreement frequently going above 30% because of persistent
stromal noise and morphological heterogeneity, as well as the
ongoing difficulty in identifying rare high-risk patterns such as
the cribriform subtype. In this study, we introduce the Hybrid
Swin-MIL Network, an automated 5-class Gleason grading (NC,
GG3, GG4, GG5 and GG4-C) framework which has been
assessed on the SICAPv2 benchmark dataset in order to address
these diagnostic problems. We have incorporated into our
pipeline an annotation-guided masking approach that
successfully shuts off the non-target background stroma so that
feature learning is focused solely on the malignant epithelial
structures while still preserving the surrounding benign
morphology. A new architecture combines an intrapatch token
attention pooling hierarchical Swin Transformer with a parallel
multi-scale CNN branch in order to capture the overall
glandular structure while at the same time keeping the fine
cellular texture. Five separate models were trained using a
patient-stratified five-fold cross-validation method and these
were combined by soft voting during inference. We achieved
95.55% accuracy, together with a 95.58% macro F1-score and
95.81% recall for the aggressive GG4-Cribriform subtype
across (12,081) previously unseen test patches using our
ensemble. Indeed, these results highlight the advantages of using
both selective input masking and dual-branch attention, and
cross-validation ensembling provides a robust, reproducible
route to fine-grained prostate cancer grading.

Beyond Reconstruction Fidelity: Task-Oriented Evaluation of Compact RGB-Derived Spectral Representations for Fruit Classification

Anchor-Preserving Coarse-to-Fine Hyperspectral Reconstruction from Mobile RGB–NIR Images

Constraint-Aware Solar MPPT Evaluation Using a Feasible-Power Reference and Pack-Terminal Energy

This paper introduces a constraint-aware evaluation framework for solar-battery MPPT. A pre-action, state-conditioned feasible-power reference distinguishes source-side tracking loss from intentional curtailment imposed by battery, protection, converter-current, and thermal limits, while pack-terminal energy measures useful delivered energy. Using equal-budget training-only tuning and independent paired uncertainty testing, 1,708 simulations of a nominal 30 W, 3S1P charger showed a positive mean IncCond-minus-P&O pack-energy difference across all four held-out profiles. The result is conditional on the modeled domain and does not claim universal superiority or hardware validation.

A Machine-Learning Surrogate Framework for Designing and Optimizing the Char Nizam Off-Grid Hybrid Power System in Bangladesh

1)A cost-minimal, reliable system design for Char Nizam: a feasible configuration (35.8 kW PV, two 10 kW turbines, battery, and diesel backup) that serves the load with < 0.1% unmet energy at an LCOE of $0.228/kWh, a 27.8% lower cost than a comparable grid search finds, demonstrably meeting objectives (i) and (ii).
2)A green, self-sufficient operating point: the design draws 50.6% of its energy from carbon-free renewables under an explicit renewable-share constraint, displacing diesel and its emissions while remaining fully islanded, meeting objectives (iii) and (iv).
3)The enabling method: a surrogate-and-search frame- work that replaces the exhaustive dispatch loop with a learned cost model. An open-source hourly simula- tor provides ground truth; a gradient-boosted surrogate predicts LCOE, renewable fraction, and feasibility in 0.13 ms (R2 ≥0.998, 36× faster than a full simulation); and differential evolution searches it, giving a measured O(g4) scaling gap and a measured 322× speedup over a g = 20 grid.
4)A justified model choice, and an audit of the framework’s own claims: we benchmark five regression families and select gradient boosting on the joint accuracy– latency criterion; separate the discretization advantage from the surrogate’s own contribution by running the same continuous search directly on the simulator, then close most of the remaining gap by adaptive refinement (7.1% → 3.4% for 100 extra simulations); cross-check the simulator against a published HOMER Pro study of this island; verify optimizer stability over 30 seeds; and decompose the LCOE band, which shows fuel price carries essentially all the risk.

Explainable Risk Detection in Blockchain Smart Contracts Using Natural Language Processing

This paper presents a lightweight machine learning framework for detecting six classes of Ethereum smart contract vulnerabilities directly from raw Solidity source code. Its main contribution is the combination of word and character-level TF-IDF features with domain-specific Solidity pattern counts, avoiding the need for bytecode compilation or graph construction. The study also uses a leakage-controlled SMOTE and vectorization pipeline, making the evaluation more methodologically reliable. Among six classifiers, LightGBM achieves the strongest performance, reaching 97.74% test accuracy and 97.07% cross-validated accuracy. A further contribution is the SHAP analysis, which connects influential model features to recognizable constructs such as .call(), delegatecall, block.timestamp, and block.number, making the predictions more interpretable for smart contract security analysis.

Evaluation Blind Spots in IEC 60870-5-104 Intrusion Detection: Beyond Attack-Only Metrics

We demonstrate that attack-only metrics under family-disjoint evaluation of IEC 60870-5-104 IDS can conceal significant false-positive rates: a 40% benign holdout reveals 5.8–10.5% FPR despite attack F1 > 0.99. We further show that 48–50% of false positives are shared across RF, XGB, and CNN, suggesting these errors stem from feature-space difficulty rather than classifier-specific overfitting.

Machine Learning for Literacy Rate Prediction in Bangladesh: Leveraging Engineered Socio-Economic Features and Explainable AI

This study presents a machine learning framework for predicting and analyzing literacy rates in Bangladesh, using a wide range of socio-economic and demographic data. The goal is to provide a practical statistical tool that helps explain the complex factors shaping national literacy. To gather useful information from the data such as GPD per capita and population density, we applied data preprocessing and feature engineering. We tested several regression models and among them the Random Forest regression model performed best achieving high prediction accuracy with a $R^{2}$ score of 0.992. A major contribution of this study is the use of Explainable AI (XAI) with SHapley Additive exPlanations (SHAP). This allowed us not only to predict literacy rates but also to understand how different factors influence the model’s results. The SHAP analysis showed that population patterns and economic indicators such as population density and gross national income (GNI) are the most important drivers of literacy outcomes. In summary, this study offers a highly accurate predictive model and also explains the complex relationships between different variables making it a valuable resource for both researchers and policymakers.

DART: An Analytical Failure-Attribution Study for Multimodal Edge Assistants

The significant contribution of this research is the development of DART, a five-layer Deployment-Aware Reliability Taxonomy for identifying where failures occur in multimodal edge assistants. It separates task-path failures from runtime/resource and connectivity-related exposure, enabling more precise reliability diagnosis than a single overall success rate. The framework is further supported by layer-specific metrics and a real case-study analysis using NORA.

Seeing Beyond the Fake: Detecting Deepfakes Using Deep Learning-Based Computer Vision

The main advantage of this research is the creation of the lightweight, leakage-aware spatial-temporal framework for deepfake detection. The work proposes a custom CNN with channel-wise attention, batch normalization, dropout and L_2 regularization, with approximately 511K trainable parameters. It also introduces frame-level detection to video-level analysis with an Attention-BiLSTM and fusion-based classification approach. One methodological contribution is the video-first data splitting to avoid data leakage by keeping all of the frames from the same video in either the training set or the test set. The study also includes a comparison of the spatial, temporal and fusion approaches, where the fusion approach attained 93.50% accuracy at the video level and 97.99% AUC.

SwinMLP-XAI: An Explainable Swin Transformer-Based Framework for Automated Detection of Tuberculosis in Chest Radiographs

This study introduces an explainable deep learning framework (SwinMLP-XAI) using a Swin-Tiny Transformer backbone to automate tuberculosis detection. It overcomes existing internal validation and interpretability issues by incorporating Integrated Gradients for clear attribution mapping, achieving 99.84% classification accuracy.

Beyond Image-Level Random Splits: A Leakage-Aware Reliability Assessment for Lightweight Jackfruit Leaf Health-State Recognition

This study presents a leakage-aware reliability evaluation framework for jackfruit leaf health-state recognition, addressing the risk of inflated performance estimates caused by image-level random splitting. It investigates how visually similar or related images distributed across training, validation, and test partitions can introduce hidden data leakage and compromise the reliability of deep learning evaluation.

To mitigate this issue, a conservative leakage-control visual grouping (LCVG) strategy is developed to enforce separation of visually related samples before model assessment. The proposed LCVG-safe evaluation protocol is systematically compared with conventional random splitting across multiple deep learning architectures, independent random seeds, and reconstructed dataset partitions to quantify the impact of leakage on reported performance.

Furthermore, this study provides a comprehensive reliability analysis through uncertainty estimation, calibration assessment, robustness evaluation under controlled corruptions, lightweight deployment analysis, and persistent error investigation. The findings highlight the importance of leakage-aware evaluation for developing trustworthy deep learning systems for practical agricultural applications.

Explainable 3D Swin Transformer for Tumor Mass Presence Classification and Quantitative Localization in Pancreatic CT: A Comparative Study

We compared six deep learning architectures for pancreatic tumor localization in CT imaging. Our 3D Swin Transformer achieved the best performance (93.49% accuracy, 0.973 AUROC), validated through 5-fold cross-validation and statistical testing, and produced more accurate Grad-CAM++ explainability results than a 3D ViT baseline.

An Android-Assisted Accurate Transfer Learning Framework for Alzheimer’s Disease Diagnosis

Alzheimer’s disease is a progressive and devastating neurodegenerative disease caused by the shrinkage of brain tissues, which leads to memory loss, cognitive decline, and other symptoms, and its diagnosis needs magnetic resonance imaging scans of the brain. This incurable disease can be deadly if proper treatment and lifestyle are not followed. Early and accurate diagnosis of this disease is crucial, but its magnetic resonance imaging-based traditional diagnosis is cumbersome and costly. Efficient automated diagnosis is highly required for modernizing the treatment and management of this disease. This study aims to develop an Android application based on the transfer learning approach, where the performance of EfficientNet-B3 and VGG16 were evaluated to select a effective model for accurately diagnosing four stages of Alzheimer’s disease, such as non-demented, very mild demented, mild demented, and moderate demented. The both models were utilized with a dataset of 12800 magnetic resonance imaging images through the transfer learning approach, where the EfficientNet-B3 outperformed VGG16 model and achieved 97.14% accuracy, which indicated its efficiency in diagnosing Alzheimer’s disease. After several evaluations, the EfficientNet-B3 model was integrated with an Android application for making Alzheimer’s disease diagnosis easier and cheaper than the traditional approach. Besides diagnosis, it has several features for enhancing the quality of life of patients with Alzheimer’s disease.

Kolmogorov Arnold Network for Multiple Disease Classification: Interpretability and Faithfulness Evaluation

Using five clinical datasets, this study evaluates KANs for interpretable medical classification. It compares KAN with established ML models and evaluates its intrinsic spline-based explanations against SHAP, permutation importance, and Random Forest feature importance. A deletion-based faithfulness analysis further verifies whether the identified important features genuinely influence model predictions.

An Acoustic Dataset for Deep Learning-Based Non-Destructive Testing (NDT) of Brick Quality via Mel-Spectrogram Analysis

To determine the structural integrity of clay bricks, vision-based methods face many limitations. Another conventional testing method is to assess the impact sound of bricks using human subjective auditory perception. In this context, we present a comprehensive acoustic dataset of brick impact sounds to advance machine learning-based research in objective non-destructive testing (NDT). First, audio data were collected from three active brick kilns—MTB Brick Field, Amin Brick Field, and Trishal Brick Field. For each recording case, two bricks are struck together at a distance of 20-30 cm, and the resulting sound is recorded with a BOYA BY-M1 Pro II microphone. The dataset’s ground-truth labels were determined using preselected, factory-graded brick batches. The complete dataset contains a total of 1,426 distinct impact audio samples, classified into three quality grades: Grade A (465), Grade B (509), and Grade C (452). Second, the audio samples were converted to mel-spectrograms to accurately capture time-frequency acoustic features. This process included noise reduction and silent segment pruning. Third, machine learning-based classification experiments were conducted to evaluate the dataset’s performance. A classification accuracy of 82.74\% was achieved by applying a random forest classifier to the smoothed time-frequency features. These results indicate that the acoustic features present in brick impact sound can be mathematically separated into distinct quality grades. The presented dataset can serve as a useful basis for developing deep learning-based models and for non-destructive testing (NDT) research aimed at automatically assessing the structural quality of bricks.

XAI-Droid: An Explainable Hybrid Static and Dynamic Analysis Framework for Android Adware–Trojan Malware Detection

In this paper, XAI-Droid: Explainable Hybrid Machine Learning Framework for Android Adware–Trojan Malware Detection is proposed which is a combined桐approach that explores both the static and dynamic features of APKs for detection of Android Adware–Trojan malware. To enable accurate, transparent and reproducible malware analysis, the proposed framework offers a leakage controlled learning pipeline, ensemble-based classification and SHAP inspired explainability. XAI-Droid interprets the security decisions into identifiers of influential permissions, API characteristics and runtime behaviors, which can benefit practical Android malware screening and cyber defense applications.

Robust Fabric Defect Classification Using Deep Learning with Grad-CAM Explainability

Most fabric-defect papers report accuracy from one train-test split, so there’s no way to know if it holds on a different slice of data. This work pairs five-fold stratified cross-validation with a fully separate held-out test set across five CNN architectures under identical conditions, something none of the compared prior work does. EfficientNet-B3’s 98.26% test accuracy tracked its cross-validation estimate closely, not a lucky split. Grad-CAM visualizations suggest the model attends to actual defect regions rather than background shortcuts, feeding into a practical 0.85-confidence rule for routing uncertain cases to human review.

Machine Learning Approaches to Child Malnutrition Prediction in Bangladesh: A Dual-Round BDHS Analysis with Concept Drift Detection and Web-Based Screening

Bangladesh’s child-malnutrition profile presents a paradox: stunting declined by 7.7 percentage points between BDHS 2017-18 and 2022, whereas wasting increased by 3.2 percentage points. No previous ML study has explained this divergence while also developing a deployable screening tool. This paper presents NutriGuard BD, integrating BDHS data (n = 10,705 children under five), a community survey of 52 mothers, and five key-informant interviews with paediatricians. We benchmark seven classifiers-LR, RF, GB, XGBoost, LightGBM, CatBoost, and MLP-with a novel Composite Vulnerability Index (CVI) and a leakage-free pipeline (SMOTE on training folds only, Optuna Bayesian optimisation, Youden’s J threshold tuning). CatBoost achieves the best dietary quality AUC (0.636; CV-AUC 0.632); after threshold optimisation, LightGBM achieves clinical recall of 0.769 (stunting) and 0.714 (underweight). Critically, all six non-MLP classifiers collapse to near-random wasting prediction in 2022 (AUC 0.479-0.497; 2017-18 baseline AUC=0.935)-the first architecture-independent surveillance-level evidence of concept drift in BDHS modelling. SHAP identifies wealth index (v190) as the dominant predictor (mean |ϕ| = 0.377 Spearman ρ > 0.97 cross-architecture). Iron-course completion is only 11.5%; its sole significant predictor is iron receipt in pregnancy (p = 0.018, ϕ = 0.292). NutriGuard BD delivers these models as a bilingual (English/Bangla) Django web application for community health workers with zero patient-data persistence.

Murnitin: An Explainable, Adversarially Robust and Privacy-Preserving Academic Integrity Framework for the LLM Era

Current AI detectors like Turnitin act as black boxes, frequently mislabel non-native English writing and fail against simple evasion tricks while storing student papers in private databases. In this paper, we present Murnitin, an open framework that addresses these issues directly. We combine sentence-level explainable metrics (perplexity and burstiness) with real-time defense against zero-width and homoglyph attacks. Crucially, we introduce keystroke process verification to confirm authentic human drafting, which cuts the false positive rate on non-native English essays down to 2%. Finally, we implement zero-knowledge cryptographic hashing so institutions can verify document similarity without ever keeping raw student text on central servers.

Two Public CKD Benchmarks Are Not Independent: A Reproducible Audit of Leakage, Cohort Separability, and Pseudo-External Validatio

This work provides record-level evidence that two widely-used public CKD benchmark datasets (the UCI 2015 Indian release and the 2020 Bangladeshi release) are not independent: exact-match bipartite record linkage shows all 200 patients in the 2015 dataset also appear in the 2020 dataset, contradicting their documented separate provenance. The paper demonstrates that models trained on one and “externally validated” on the other are not measuring transportability but rather leakage — near-ceiling ROC-AUC performance (up to 1.000) is shown to result from post-diagnosis outcome-correlated variables rather than genuine predictive power, dropping to a more realistic 0.909 once such variables are removed. The study contributes a reusable, dataset-agnostic auditing methodology (a provenance gate, feature taxonomy, and leakage-prevention framework) for detecting duplicate contamination and incorporation bias in clinical prediction benchmarks generally.

Ensemble Deep Learning with Test-Time Augmentation for Breast Ultrasound Classification on BreastMNIST

Our main contribution was a methodical testing of three different models of heterogeneous ensemble CNN–Transformer—ConvNeXt-Tiny, EfficientNetV2-S and Swin-Tiny—using test-time augmentation (TTA) with soft voting, all optimized for the BreastMNIST classification task. The ensemble performs with a test accuracy of 92.31%, higher than that of the best individual model, and class-wise error and inter-model agreement analyses are included to give better estimates of prediction reliability. Critically, the paper formalizes this contribution as an empirical ensemble evaluation rather than a new learning algorithm.

Lung-Ensemble: A Secure Multi-Stage Diagnostic Pipeline for Multi-Class Respiratory Disease Detection using a Stacked Ensemble and Explainable AI

Abstract—Lung diseases continue to be one of the fundamental
health threats for a significant portion of the global population.
Respiratory health and lung function, however, face severe threats
from various infections that significantly increase mortality rates.
In addition, there is a growing need for rapid detection and
accurate diagnosis of these diseases to prevent severe clinical
complications. In this paper, we have explored and analyzed
several state-of-the-art deep learning models for multi-class lung
disease detection using a comprehensive dataset of chest Xrays. From the models we tested, we noticed that MobileNetV2
and VGG16 gave the best individual results. To leverage these
strengths, we propose a secure, multi-stage diagnostic pipeline
called Lung-Ensemble. This framework utilizes a binary ”Gatekeeper” classifier to filter out non-medical images, securing
the system against erroneous processing, before passing valid
X-rays to a feature-fusion based ensemble utilizing VGG16,
DenseNet121, and MobileNetV2. Our proposed model (LungEnsemble) shows a significant improvement, achieving a final
accuracy of 98.33%.

CEEC: Mixed-Precision EfficientNetB3 Transfer Learning for Fine-Grained Endemic Bird Classification and Biodiversity Monitoring

Abstract—Identifying bird species is crucial for biodiversity
conservation, yet many of the 25 endemic Indian species are
visually similar and hard to distinguish. Fine-grained classification is challenging due to high intra-class variability in
pose, lighting, and background, coupled with subtle inter-class
differences in plumage and morphology. Despite growing use
of deep learning for avian identification, large-scale, balanced
datasets and tailored models for Indian endemic species remain
scarce. This paper presents a deep learning framework trained
on a curated dataset of 37,500 images spanning 25 endemic
Indian bird species. Our contributions include: (i) a balanced,
annotated dataset; (ii) a reproducible pipeline with systematic
data augmentation and mixed-precision training; and (iii) a
novel EfficientNetB3-based architecture (CEEC) optimized for
fine-grained classification. The CEEC model employs a frozen
EfficientNetB3 backbone pretrained on ImageNet, combined with
a custom classification head incorporating batch normalization,
dropout, and dense layers to reduce overfitting while preserving discriminative features. Among evaluated models, including
custom CNNs, ResNet152V2, and EfficientNet variants, CEEC
achieved superior performance: 98.74% test accuracy, 0.05487
test loss, and 0.9873 F1-score.

GateDistill-NPK: A Novel Reject-Aware Distillation Framework for Rice-Leaf Macronutrient-Deficiency Diagnosis

Deficiencies of nitrogen (N), phosphorus (P), and potassium (K) constitute a leading yet largely correctable cause of rice-yield loss, but operational diagnosis remains constrained by scarce agronomic expertise and by destructive laboratory assays that do not scale to smallholder cultivation. Prevailing image classifiers entangle two decisions that ought to be separated, namely whether a leaf is a deficient paddy leaf and, conditionally, which nutrient is depleted, and they provide no mechanism for rejecting the out-of-domain imagery that field acquisition unavoidably yields. This work introduces GateDistill-NPK, a reject-aware cascade that decouples detection from differential diagnosis. A first-stage EfficientNet-B0 gate assigns each image to a deficient, healthy, or non-paddy outcome, converting out-ofdomain inputs into an explicit rejection rather than a spurious nutrient label. Only deficient images advance to a compact Scratch-Fusion MobileNetV2 student (SF-MV2), which fuses a pre-trained MobileNetV2 backbone with a from-scratch convolutional branch and is optimized by temperature-scaled distillation from a ResNet-50 teacher; a deterministic rule then maps the prediction to a fertilizer prescription. Evaluation uses a threesource corpus of 8,037 images, deliberately assembled so that each gate category originates from an independent acquisition domain. The gate attains 94.35% test accuracy with a 0.9732 micro-averaged area under the curve (AUC), while SF-MV2 delivers 87.83% accuracy, 0.9181 balanced accuracy, and a 0.9789 AUC using only 2.72 M parameters, an 8.8× compression that preserves accuracy comparable to its teacher. These results establish that reject-aware gating and cross-capacity distillation can be unified within one edge deployable pipeline for trustworthy in-field nutrient diagnosis.

Secure Federated Learning for Brain Tumor Segmentation Under Non-IID and Missing-Modality Conditions

This paper presents a lightweight (<1M-parameter) federated 3D U-Net for multi-region (WT/TC/ET) brain tumor segmentation under realistic hospital conditions — non-IID client data and a per-client missing MRI modality — evaluated with a modality-aware Dice/HD95 protocol. We show that ignoring a client's missing modality at evaluation time inflates pooled Dice by 0.075–0.12 points, a bias not disclosed in prior federated segmentation work. We also validate a full secure-deployment stack (TLS, ECDSA node authentication, SecAgg+) on real gRPC connections, and demonstrate that a gradient-inversion attack recovering a 15.51 dB PSNR patch from a raw update drops to −66.03 dB once SecAgg+ masking is applied — an 81 dB gap quantifying secure aggregation's practical privacy benefit.

The Evolution of Cyber Attack Vectors: From Network Exploits to Autonomous AI Threats

1. A unified taxonomy of cyber attack vectors covering network, cloud, IoT, CPS, human-centric, and AI/autonomous-agent attacks.
2. A cross-domain comparative analysis highlighting how cyber attacks have evolved from isolated vulnerabilities to interconnected and intelligent attack scenarios.
3. A research roadmap identifying key challenges and future directions for securing AI-driven and autonomous digital ecosystems.

A Graph-Based Framework with Patient Similarity Clustering, Imbalance-Aware Learning, and Multi-Level Explainability for Reliable Stroke-Risk Prediction

This research presents a reliable and interpretable framework for stroke-risk prediction by combining patient similarity graphs, clustering, graph attention networks, and imbalance-aware learning. Patients are represented as nodes in a mutual k-nearest-neighbor similarity graph, while Ward-linkage agglomerative clustering is used to capture community-level risk patterns. A two-layer Graph Attention Network with focal loss and an F2-based decision threshold is then used to improve the detection of stroke-positive patients, with particular emphasis on reducing false negatives. The framework also integrates GNNExplainer, SHAP, and counterfactual analysis to provide explanations at the patient, neighbor, and cluster levels. Experimental results show that the proposed GAT with agglomerative clustering provides stronger recall-oriented performance than the evaluated alternatives, while cross-validation, cluster stability, and explanation-consistency analyses provide additional evidence of reliability.

UniPedFormer: A Unified Single-Stage Vision Transformer for Joint Pedestrian Detection and Trajectory Prediction

We describe a monocular, pedestrian-specific single-stage architecture in which one transformer encoder serves both detection and short-horizon trajectory prediction, and we position it explicitly against the existing joint perception-and-prediction literature rather than around it.

We introduce a Spatial-Temporal Decoder that composes pedestrian self-attention, scene cross-attention, and temporal self-attention inside each decoder layer, so that crowd context, visual grounding, and motion history are mixed before either head reads the query.

We show that a four-term objective combining focal, smooth-$L_1$, ADE, and FDE losses converges without the loss balancing problems that often affect multi-task training, and we report the resulting behaviour honestly, including the synthetic nature of the benchmark and the metrics it does not support.

SimSighs: Benchmarking and Explaining Transformer Models for Bengali Telecom Complaints

The increasing amount of customer feedback in
Bengali presents a significant problem for telecom operators.
Analyzing and classifying such data takes a lot of time and does
not always produce accurate results. This article introduces
SimSighs, an improved and relabeled benchmark dataset for
Bengali telecom complaint classification. We aim to address the
problems caused by inconsistencies in existing public datasets,
merge classes with similar meanings, and balance the number of
samples across all classes. We also evaluate twelve different
models, ranging from traditional machine learning to deep
learning and state-of-the-art transformer architectures. Our
experiments demonstrate that models based on transformer
architecture, especially Bangla-BERT, achieve the best results
in terms of F1-score, reaching 92.22%. We apply Explainable
AI (XAI) to understand which parts of the input text contribute
most to a model’s decision. The visualization technique, Layer
Integrated Gradients, helps identify patterns and words that a
model uses to classify a complaint based on its category. Overall,
our research emphasizes the importance of high-quality datasets
and the application of XAI methods for developing dependable
and accurate NLP systems for customer feedback analysis.

Electromagnetic Transient Modeling and Simulation of Inverter-Based Black Start: A Comprehensive Review of Grid-Forming Control, Current Limiting, and System Restoration Dynamics

Abstract—The accelerating integration of inverter-based re
sources (IBRs)—including battery energy storage systems
(BESS), photovoltaics (PV), and variable-speed wind turbines—
into modern power systems has created a compelling impetus
to re-examine conventional black-start and system restoration
paradigms. Traditional black-start capability has been exclusively
provided by synchronous generator-based power plants, whose
inherent short-circuit current capacity (6–8 p.u.), electromagnetic
inertia, and well-characterised electromechanical dynamics facil
itated robust sequential network energisation. IBRs, by contrast,
present sharply limited overcurrent capability (typically 1.1
1.5 p.u.), possess no physical inertia, and require advanced grid
forming (GFM) control architectures to autonomously establish
and sustain terminal voltage on a de-energised network. This
paper provides a systematic and critical review of electro
magnetic transient (EMT) modelling frameworks, simulation
methodologies, mathematical control formulations, and transient
phenomena pertinent to IBR-driven black start. Key topics
include droop-based GFM control with integrated negative
sequence voltage compensation, phase-by-phase current limiting
in the abc frame, transformer inrush mitigation via voltage-ramp
soft-start, induction motor starting dynamics under constrained
IBR current, collective parallel GFM operation without leader
follower communication, GFL DER cooperative restoration, and
VSC-HVDC black-start control modes.
Index Terms—Black start, electromagnetic transient (EMT)
simulation, grid-forming (GFM) inverter, inverter-based re
sources (IBR), negative-sequence control, phase current limiter,
droop control, VSC-HVDC, transformer inrush, induction motor
starting, system restoration, microgrid.

A Comparative Study of Machine Learning Algorithms for Nationwide Air Quality Index Prediction in Bangladesh

The contribution of this paper to the research is that it provides a nationwide and multiregion framework for AQI forecasting in Bangladesh, filling in a critical gap within the literature. In particular:

1. Geographic scale – While previous research done in Bangladesh only used a few cities or regions for AQI prediction, this study has utilized a dataset consisting of 103 cities/regions for approximately 25 years (2000-2025).
2. Chronological split for time-series validity – This framework has utilized a chronological and region-wise train-test split and region-wise lag feature engineering as opposed to randomized split and non-region-wise lag feature engineering, providing a better simulation of forecasting.
3. Comparative evaluation between models – This framework sets a benchmark for evaluating the models in a consistent environment using the same evaluation metrics (RMSE, MAE, MAPE, and R²) at a national level and at a regional level explicitly.
4. Flexible framework – It provides foundations for future implementation of XGBoost, LightGBM, GRU, Transformer, graph models, and even IoT-based real-time implementation.

Overall, the paper specifically says that the experimental results of the study are illustrative, not measurements obtained during the model training process. Therefore, the actual contribution of this work is the framework and methodology themselves.

Ink Meets Intelligence: A Computer Vision Based Approach for Bengali Handwritten Character Recognition

This research presents a practical and lightweight approach to Bengali handwritten character recognition. We developed an 84-class dataset containing 168,528 images of vowels, consonants, and compound characters by combining newly collected handwriting samples with the BanglaLekha-Isolated dataset. We also compared a custom CNN-MobileViT model with an adapted MobileNetV3-Small model. The best model achieved 94.71% test accuracy while using only 1.12 million parameters, making it suitable for devices with limited computational resources. In addition, we introduced a confusion-aware loss to give greater attention to visually similar characters and used Grad-CAM and LIME to explain which parts of each character influenced the model’s predictions.

A Spatio-Temporal Deep Learning Approach for Edge Deployable Camera-Based Detection of Cardiac Arrest Symptoms Using Generative-AI Based Synthetic Data

This research develops an edge-deployable camera-based system for detecting visible cardiac-arrest-related symptoms using Generative-AI-based synthetic video data and limited real recordings. Four spatio-temporal deep learning models were comparatively evaluated, achieving up to 98.11% accuracy, while a lightweight CNN-LSTM model was deployed on a Raspberry Pi 4 for real-time symptom monitoring and alert generation.

e, Who Must Not Be Named: Cryptanalysis of LCG-Based Randomness in Guillou-Quisquater

This work presents a lattice and polynomial-based attack on the Guillou–Quisquater identification protocol when its ephemeral randomizers are generated by a known-parameter LCG. The attack exploits short integer relations among challenges to eliminate the secret-dependent term, recover the initial LCG state through polynomial GCD computation, and subsequently recover the secret, without requiring the public exponent(e) during the recovery procedure.

Explainable Machine Learning Framework for Depression Risk Prediction

Depression is a significant mental health issue and early recognition of those who are vulnerable may aid in timely intervention. In this study, an explainable machine learning framework for predicting depression risk based on demographic, socioeconomic, behavioral, clinical, examination and laboratory variables is presented. The data was also preprocessed and selected using feature selection techniques and class imbalance was corrected from the training set with SMOTEENN. Nine machine learning models were evaluated, including Logistic Regression, Decision Tree, Random Forest, K-Nearest Neighbors, Support Vector Machine, Naïve Bayes, Extra Trees, Gradient Boosting, and XGBoost. Overall, XGBoost performed best, with an accuracy of 82.60% and a ROC-AUC of 0.760. To make the model more interpretable and to find influential factors related to the depression risk, SHapley Additive exPlanations (SHAP) was used. The framework offers global and local explanations, making predictions more transparent. The results suggest that effective predictive modeling coupled with EAI can be used to facilitate a more reliable and interpretable evaluation of depression risk and could be part of future healthcare decision-support systems.

DIAsm: A Bilingual AI-Assisted Mobile Health Application for Type 2 Diabetes Self-Management in Bangladesh

This research presents DIAsm, a deployed bilingual English–Bangla Android application that integrates multiple Type 2 diabetes self-management functions within a single platform designed for the Bangladeshi context. Its main contribution is the combination of health monitoring, reminders, lifestyle and nutritional support, diabetes education, risk assessment, wellness utilities, locally relevant food information, and a grounded conversational assistant. The assistant employs a hybrid RAG pipeline incorporating rule-based intent classification, multilingual semantic retrieval, metadata filtering, heuristic reranking, and constrained local LLM-based response rewriting. Evaluation included a 50-query technical benchmark and a post-deployment study with 65 users, demonstrating the feasibility and positive perceived usability of this integrated approach.

Multivariate Modeling of Employee Retention Using Reward, Work-Life Balance, and Work Environment Factors in Malaysia’s Industrial Equipment Sector

Employee retention has become a strategic necessity in human resource management, specifically in the engineering and industrial distribution sectors. Specialized skills are rare and essential for operational continuity, customer service reliability, and profitability. The purpose of this study is to explore the
empirical factors that affect employees’ exit from Sime Darby Industrial (SDI) Sdn Bhd, a franchisee distributor of Caterpillar heavy machinery in Malaysia, with a high rate of employees’ voluntary turnover that has been observed to be over 40% for the
past decade. Based on the three theories, a quantitative deductive investigation was performed by administering a 35-item survey to N = 60 mid-level technical executives, engineers, diagnostic specialists, and managers (with a 50% return rate) from a target population of N = 120 that represents the extent of the sample. The extent of the sample to be used in the investigation was N = 120, but a 50% return rate of N = 60 mid-level technical executives, engineers, diagnostic specialists, and managers was achieved for a quantitative deductive investigation, founded on the three theories. Construct reliability (α), normality diagnostics (skewness/kurtosis z-scores), Pearson correlation, and Ordinary Least Squares (OLS) multiple linear regression were used for the analysis of the empirical data in IBM SPSS 23. The results show that reward and recognition have a statistically significant positive direct effect on employee retention (r = 0.323, p = 0.012; regression β = 0.425, t = 2.620, p = 0.011). By contrast, both Work-Life Balance (r = 0.173, p = 0.186) and Work Environment (r = 0.042, p = 0.750) have weak direct bivariate impacts but
strong, statistically significant bivariate correlations with the composite intervening psychological construct (Job Satisfaction, Organizational Commitment, and Employee Loyalty) (r = 0.546 and r = 0.658, respectively; all p ¡ 0.001). The intervening construct, in turn, significantly contributes to retention (r = 0.311,
p = 0.016). The overall regression model is statistically significant (F(3, 56) = 3.109, p = 0.034, R² = 0.143). Proposed to create a strategic retention architecture in 5 pillars for the industrial machinery businesses.

User Interface Evolution of YouTube (2005-2026): A Longitudinal Analysis from HCI Perspective

YouTube has evolved from a simple video-sharing website into a global social-media platform used by billions of viewers across diverse languages, abilities, and levels of digital experience. This study examines how its user interface changed from 2005 to July 2026 and what those changes imply for usability and user control. A longitudinal archival dataset of 22 YouTube homepage snapshots was collected from the Wayback Machine. Semiotic analysis traced changes in navigation, thumbnails, feedback mechanisms, account identity, and search, while a usability evaluation rated seven archive-visible dimensions on a 0-5 ordinal scale at representative checkpoints. Supporting structural properties of the archived HTML were also examined to identify implementation-level changes. The results identify five broad time periods: YouTube moved from explicit, text-heavy navigation and community ratings toward visual discovery, persistent identity, personalization, recommendation-driven feeds, and AI-assisted interaction. The descriptive HCI score increased from 2.86 in 2005 to 4.43 in July 2026, mainly through stronger consistency, efficiency, feedback, and accessibility support; user control improved more modestly as content discovery became increasingly algorithmically mediated. The findings suggest that successful interface evolution depends not only on adding functionality but also on preserving understandable signifiers, predictable interaction, and meaningful user agency.

An Autonomous Detection Approach for Isolated Drones of the Swarm in 3D Environments

. Swarm drones generally used to perform a specific task on Target of Inter-est(ToI). Any members(drones) of the swarm can be outlying at any time during execution time of its assigned task. If any drone outlays during the execution time, performance of the swarm may be degraded and goal may not be achieved. For this reason, detection of isolated drone is very much important. In this paper we propose an autonomous framework to detecting these isolated drones in 3D environment. At first, we collect boundary drones using Gift Wrapping Algo-rithm. But all boundary drone may not be connected at least one hop neighbor. To overcome this limitation, we offered an algorithm. This algorithm will restrict the connectivity between two boundary drone with their communication range and produce a concave polyhedron. Then we tested very drone autonomously wheth-er they lay within the area produced by concave polyhedron. If they lay in the polyhedron, it is treated as in well topology. otherwise treated as an isolated drone. We evaluate our experiment in python language in google colab environ-ment and got 100% detection accuracy.

CoopShield: An IoT-Based Multimodal Environmental Monitoring and Safety System for Smart Poultry Farming

The main research and findings of CoopShield are connected with the creation of an inexpensive and low-latency multimodal automation system, which is capable of reducing risks for the poultry farming site in real time. The device can regulate the temperature, humidity, ammonia gas level, number of PPM in the air, and flock commotion noise as a sign of stress and bad conditions. It uses a single microcontroller (ESP32) to process this information and perform relay operations for the actuators. The main unique aspect of the developed solution is the implementation of a closed-loop system with events triggered in real-time for each particular actuator to reduce the danger for the flock. Additionally, the data from sensors are continually sent through Bluetooth to Blynk for plotting and additional analysis in the cloud. Thus, the created device is an easy-to-use smart technology for farming that provides safety for animals and automation of the process.

Above Chance, Below the Baseline: A Majority-Anchored, Case-Grouped Ablation for Thyroid Ultrasound Segmentation and Report-Derived Suspicion Classification

Thyroid ultrasound computer-aided diagnosis is routinely reported as a segmentation Dice score and a classification F1 score, with little scrutiny of what drives either number on small, imbalanced, multi-image-per-patient benchmarks. We supply that scrutiny through two matched-protocol ablations on TN3K (segmentation) and DDTI (classification) under one Holm–Bonferroni-corrected statistical protocol. On TN3K, three ImageNet-pretrained encoder–decoder architectures (U-Net, U-Net++ and FPN, ResNet-34 encoder) exceed a from-scratch U-Net by 0.0670–0.0758 Dice under 3-fold cross-validation (adjusted p ≤ 0.0054), while the pretrained architectures separate only partially from one another: U-Net++ exceeds FPN (adjusted p = 0.024) but not U-Net (adjusted p = 0.103). A multi-axis ablation confirms pretraining (−0.0858 Dice) and augmentation (−0.0228 Dice) matter (adjusted p = 0.039 each); loss form and resolution do not reach significance. On DDTI, a deterministic majority-class predictor is a mandatory row in every table: all four backbones (ResNet50, DenseNet121, EfficientNet-B0, ViT-B/16) score below it on F1 (best 0.8896 vs. 0.9035) and accuracy (best 0.8162 vs. 0.8246), yet all exceed it on Matthews correlation coefficient (0.1541–0.3072 vs. 0.0000) and balanced accuracy (0.5660–0.6506 vs. 0.5000); no pairwise difference survives correction (best raw p = 0.040, adjusted p = 0.240). Replacing a naive image-level split with a case-grouped one moves F1 by −0.1439 but Matthews correlation coefficient by only −0.0143 and AUC by +0.0029. Re-running two identical configurations gives repeat-run spreads of 0.0494 Matthews correlation coefficient, 0.2217 F1 and 0.0420 Dice, larger than most measured effects. We further disclose a pixel-level ROC-AUC of 0.9804 against a Dice of 0.7568, and a Youden threshold saturating at 1.0000 with expected calibration error 0.1324.

A Trend-Based IoT Monitoring and Automated Cleaning Framework for Solar PV Systems Under Dust and Salinity Conditions

Dust, humidity, and salt deposition are some of the challenges faced by PV systems in coastal areas that can impact the output of the electrical system and make them more prone to maintenance. Typical maintenance methods are guided by manual cleaning or by monitoring a single parameter only and can lead to unnecessary cleaning and/or delayed maintenance. A smart PV monitoring and automated cleaning system for coastal area based on Internet of Things (IoT) is proposed in this paper. The proposed system includes voltage, current, dust and salinity sensors, an autonomous solar-panel cleaning robot and an ESP32 controller. The Trend Based Cleaning Algorithm compares the PV voltage it expects, the measured PV voltage, a performance loss per unit time, salinity level, and the time since the last cleaning operation. Cleaning is started only when the pre-set performance-loss and salinity requirements are met. A 7 day simulation and hardware experiments were performed. The salinity losses were from 2.1% to 5.0% and the performance losses were 4.2% to 9.8% measured in the hardware. The results of the simulation and hardware showed to be in good agreement the trends, differences were due to the environmental conditions, sensor tolerances and measurement noise. The automated cleaning cycle was about 8-12 minutes, which is less time than manual cleaning takes, which is reported as 30-40 minutes; the project reports a 70% reduction in maintenance labour. The findings show that such a comprehensive system (environmental sensing, electrical performance analysis, IoT monitoring and autonomous cleaning) for PV maintenance in a challenging coastal environment is feasible.

LLM-Assisted Service Migration and Resource Allocation for Dynamic Cloud-Edge Computing

Dynamic cloud–edge computing environments require continuous service placement and resource allocation under changing workloads, user mobility, network conditions, and heterogeneous computing resources. Conventional heuristic approaches provide efficient decision making but may lack adaptability, while deep reinforcement learning (DRL) approaches can require substantial training and computational overhead. This paper proposes an \textit{LLM-Hybrid Service Migration and Resource Allocation (SMRA)} framework that combines Large Language Model (LLM)-based candidate generation with heuristic optimization and multi-objective cost evaluation. The framework represents the current system state using workload, latency, available resources, energy consumption, bandwidth, and migration-related information, which are provided to the LLM to generate candidate migration strategies. In parallel, a heuristic module generates analytically derived candidate strategies. The feasible candidates from both modules are evaluated using a unified cost function incorporating service latency, energy consumption, and migration overhead, and the minimum-cost strategy is selected for execution. A simulation-based evaluation involving mobile users and edge servers compares the proposed framework with heuristic, DRL-based, and random baseline approaches. The reported results show reductions of approximately 29% in average service latency, 17% in energy consumption, and 22% in total system cost compared with the heuristic baseline. The findings demonstrate the potential of combining LLM-based contextual candidate generation with deterministic optimization for adaptive service management in dynamic cloud-edge environments.

DriftErase-HAR: Ledger-Anchored Federated Unlearning for Multimodal Human Activity Recognition under Asynchronous Sensor Drift

The paper proposes DriftErase-HAR, a novel federated unlearning framework designed to address the intertwined challenges of sensor drift and data deletion in multimodal human activity recognition. The framework introduces a compact, drift-adaptive backbone using mask-aware multimodal encoding, gated low-rank adapters, and coverage-aware attention-based fusion. It further presents a three-stage unlearning mechanism combining partition-weighted ledger rollback, distribution-matching data scrubbing, and fairness-aware model repair, supported by a hash-chained update ledger for auditability. Experiments on the WISDM dataset demonstrate that DriftErase-HAR substantially reduces membership-inference leakage while improving recognition accuracy and preserving client-level fairness. Moreover, it achieves near-identical parameterization to the retraining oracle with 22.79× fewer optimizer steps, demonstrating an effective trade-off between privacy, utility, fairness, auditability, and computational efficiency.

The Mediating Role of Perceived Dealer Pricing Fairness in the Relationship Between Technology Readiness and Solar Home System Adoption: Evidence from Rural Communities

This study provides strong empirical evidence from rural communities in Hatiya, Noakhali, Bangladesh, demonstrating that Optimism (OPT) and Innovativeness (INN) significantly and positively influence Solar Home System Adoption Intention (SHA), whereas Discomfort (DIS) and Insecurity (INS) do not exert significant effects. Importantly, Perceived Dealer Pricing Fairness (PDP) significantly and positively influences SHA and mediates the relationships between OPT and SHA and between INN and SHA. These findings establish fair dealer pricing as a critical market-based mechanism through which positive technology-related dispositions can translate into stronger SHS adoption intention, offering a context-specific contribution to renewable-energy adoption research in geographically remote rural communities.

Digital Twin-Driven Framework for Predictive Cyberattack Detection and Autonomous Resilience in IoT-Enabled Smart Grids

This research proposes a Digital Twin-driven framework for predictive cyberattack detection and autonomous resilience in IoT-enabled smart grids. The key contribution is the integration of real-time Digital Twin modeling, AI-based cyberattack prediction, and autonomous response mechanisms to continuously monitor grid behavior, identify anomalous activities, and mitigate threats before they cause significant disruption. The framework enables adaptive threat intelligence, real-time state synchronization, and self-healing responses, improving the security, resilience, reliability, and operational continuity of smart-grid infrastructures against evolving cyberattacks.

RETINO GUARDAI:ANEFFICIENTNETMODELANDDEEPLEARNING FRAMEWORKFOR EXPLAINABLE,ANATOMICALLY-AWAREDIABETIC RETINOPATHY SCREENING

Diabetic Retinopathy (DR) remains one of the lead
ing preventable causes of vision loss among working-age adults
with diabetes, and international screening guidelines recommend
regular fundus examination to catch referable disease before
irreversible damage occurs. Manual grading, however, is slow,
requires a trained ophthalmologist, and does not scale to the
size of most diabetic populations, which motivates automated
screening from retinal fundus photographs. This work presents
RetinoGuard AI, an integrated screening pipeline built around a
fine-tuned EfficientNet-B4 backbone for binary DR classification,
reaching 91.4% validation accuracy on the APTOS 2019 Blind
ness Detection dataset with a sensitivity of 92.1% and specificity
of 90.8% on referable cases. Beyond classification, the system
contributes two further components that most published DR
classifiers do not provide: first, a Grad-CAM based explainability
module that produces a pixel-level saliency map showing exactly
which parts of the fundus image drove the prediction; and
second, a Five-Zone Retinal Region Activation Analysis that
aggregates that saliency map over five anatomically named
zones- the macula, optic disc, superior arcade, inferior arcade,
and periphery- condensing a dense heatmap into a small
set of clinically meaningful numbers a reviewer can scan in
seconds. All three components are wired into a single session
authenticated Streamlit web application, so that classification,
visual explanation, and zone-level risk summary are produced
end-to-end from one uploaded image, without a separate offline
analysis step.

A Comparative Study of Spatial, Frequency-Domain, and Color-Channel Preprocessing Filters for Fundus-Based Cataract Detection

Evaluation of Significance of Rechargeable Battery with Special Emphasis on Lithium-ion Technology

This research provides a comprehensive evaluation of rechargeable battery technologies, with particular emphasis on lithium-ion batteries, covering their evolution, operating principles, performance characteristics, applications, challenges, and future prospects. It highlights the critical role of lithium-ion technology in renewable energy integration, portable electronic devices, electric vehicles, and other emerging applications, while identifying key challenges related to safety, cost, raw materials, environmental impact, and sustainability. The study also emphasizes the potential of AI and advanced materials to accelerate the development of safer, more efficient, and sustainable next-generation battery technologies

Ensemble Deep Learning Framework for Calibrated and Explainable Diabetic Retinopathy Grading on RetinaMNIST

The importance of this study is that it aims to shift the focus from achieving grading accuracy in DR towards a more reliable, clinically meaningful and interpretable grading system. The proposed four-model ensemble obtained a QWK score of 0.813 and macro-AUC of 0.908, which are better than the previously reported baseline results on RetinaMNIST and achieved better ordinal and discriminative results.

Most importantly, the study merges the ordinal-aware evaluation with the explainability, calibration, uncertainty analysis and structured error analysis of Grad-CAM. This not only provides an assessment of the accuracy of the model, but also the validity of the level of confidence, whether the model targets clinically relevant retinal lesions, and where potentially harmful errors (e.g. under-grading PDR) may arise.

Thus, the key contribution is a more holistic perspective on the assessment of AI-based DR grading, focusing not just on accuracy, but also on clinical trust and risk awareness.

Rethinking Noise Shaping in Differentially Private Federated LoRA: Why Isotropic Perturbation Wins.

Differentially private federated fine-tuning with
LoRA typically perturbs adapter updates as homogeneous vectors, overlooking the asymmetric effects of its low-rank factors.
Because the effective update is (α/r)BA, perturbations in A
and B are amplified according to the norm of the complementary factor. We formalize this asymmetry through a block-wise
Gaussian mechanism with separate clipping thresholds and noise
scales, subject to ∑︁g Cg2/s2 g = 1/σ2, which preserves the privacy
guarantee of isotropic noise. Although the proposed shaping reduces modeled noise energy, experiments show that it consistently
degrades utility. We identify three causes: degenerate single round error modeling under standard LoRA initialization, crossround cancellation between perturbation and gradient repair,
and update-direction distortion introduced by per-block clipping.
Across twelve paired experiments at identical (ϵ, δ), isotropic
noise consistently outperforms shaping, with gains of 3.02 and
7.28 percentage points on attention and feed-forward projections,
respectively. These results support isotropic perturbation as the
stronger baseline for private federated LoRA.

A Coupled Index Machine Learning Optimization Framework for River Water Quality Assessment and Control: The Halda River, Bangladesh

The maintenance of a river suitable for fish spawning demands three separate abilities that need to be separately developed: a defendable description of water quality in terms of a scalar parameter, a technique for its economical and continuous estimation, and a rule of converting it into action. This paper integrates the three into a single framework and applies it to the case of the Halda River, the only remaining natural spawning ground for Indian major carps in Bangladesh. The formula of a weighted arithmetic water quality index (WQI) is modified so that the subindex for DO is normalized by the saturation value depending on temperature; the principal term of the resulting index is then equivalent, up to rescaling, to the deficit variable of a Streeter–Phelps oxygen balance, making the index and the load allocation model commensurate in terms of the state variable. Applying the WQI to 18 station-seasons at six stations and three seasons, one obtains: the lower tidal reach at Mohora scores 92.4 (Very poor) during monsoon season and 109.3 (Unsuitable) after monsoon, the impounded reach above the Bhujpur rubber dam remains Excellent (6.3–12.9), and the seasonal average value decreases from 63.6 in monsoon to 34.3 in winter. Seven predictor variables are then learned for predicting the index value without oxygen data using four simple inputs. Extreme randomised tree model with leave-one-out cross validation achieves R² = 0.715 and RMSE = 14.66 index units, logistic regression classifier classifies acceptable and Poor-or-worse water quality states with 100% cross validated accuracy, and the predictive estimator recovers 87.1% of the index variation with 12 out of 13 alert months identified in a 36-month synthetic deployment scenario. The 5000-sample Monte-Carlo simulation analysis shows that the assimilative capacity of Mohora exceed with probability 1.00 in both wet seasons, which requires a 27–43% load reduction. An analytical solution demonstrates that the required reduction is independent of the reaeration-to-deoxygenation rate ratio. Constrained optimisation of aeration and sediment control gives the minimum cost action for Excellent class, and the oxygen rather than the turbidity becomes the binding control.

Explainable Hybrid Feature Fusion Network for Automated Skin Cancer Detection from Dermoscopic Images

A fair and explainable hybrid framework integrating EfficientNetV2L deep image features, handcrafted computer
vision descriptors, and patient metadata for multi-class
skin lesion classification.
• A preprocessing pipeline incorporating hair artifact removal and data augmentation to improve image quality
and address class imbalance.
• Integration of LIME and SHAP for local and global
interpretation of model predictions.
• Fairness-aware evaluation across estimated skin-tone
groups.

Few-Shot Hybrid Time Series and Large Language Model Framework for Air Leak Fault Diagnosis in Pressurized Pneumatic Systems

This study presents a novel few-shot multimodal framework for air-leak fault diagnosis in pressurized pneumatic systems, integrating learned multivariate time-series representations with large language model (LLM) based reasoning. The key contribution is the coupling of a convolutional autoencoder Prototypical Network with instruction-tuned LLMs through fixed and confidence weighted probability fusion, enabling complementary temporal and statistical information to be exploited under limited labeled data. The framework was validated on real-world MetroPT-3 compressor telemetry and achieved 99.0% accuracy, 99.1% precision, 99.0% F1-score, and 1.000 AUC with the Phi-3.5-mini hybrid. Moreover, confidence weighted fusion improved probability estimation, demonstrating that the proposed integration can enhance both fault discrimination and prediction reliability in few-shot industrial diagnosis.

A Multimodal Framework for Bangla Fake News Classification through Cross-Attention Fusion

We have developed a multimodal framework that integrates BanglaBERT, CLIP, AI-generated image detection,
and metadata through cross-attention fusion for Bangla
fake news detection.
• We have introduced a structured credibility mechanism
combining headline-body consistency, source credibility,
and virality features for improved fake news detection.
• We have evaluated the framework on 3,601 Bengali news
samples, including a held-out set of 491 unseen articles
to assess generalization.

An Explainable AI-Driven Deep Adaptive Learning Framework for Breast Cancer Classification Using Histopathology Images

A unified pipeline benchmarking three CNN backbones
under Plain and Transfer Learning regimes on BreakHis,
identifying fine-tuned DenseNet121 as the most accurate
configuration.
• Grad-CAM explainability integrated on the best model,
enabling inspection of the tissue regions driving each
prediction.
• An Adaptive Learning Engine that updates the deployed
model incrementally using a replay buffer and a safetyrollback check, avoiding both full retraining and catastrophic forgetting.
• An end-to-end incremental-update case study demonstrating the engine, including its versioning and audit-logging
mechanism, and a comparison against ex

An Explainable AI Approach for Bus Safety Risk Classification in Data-Scarce Transport Networks

Bus overtaking and reckless driving are important
road-safety concerns in data-scarce low- and middle-income
cities. This paper develops a machine-learning and explainable-AI
(XAI) framework for Dhaka, Bangladesh, where structured bus
telematics, digital driver logs, mechanical-condition histories, and
bus-specific incident labels are not readily available. The central
contribution is the Dhaka Bus Driving Risk Dataset (DBDRD), a
domain-grounded synthetic benchmark containing 12,000 records
and 43 variables across six feature groups. Its generation
process combines context-specific distributions, explicit conditional
dependencies, a 12-component literature-calibrated risk score,
Gaussian perturbation, percentile-based ordinal labeling, and
controlled missing-at-random noise. Random Forest, XGBoost,
LightGBM, and radial-basis-function SVM were evaluated using
a leakage-controlled 70/15/15 split and cross-validated tuning.
XGBoost achieved the strongest overall test performance with
76.4% accuracy, 0.761 macro F1, 0.767 weighted F1, and 0.937
one-versus-rest ROC-AUC. SHAP analysis identified operational
risk, aggression, compliance risk, and speed excess as the dominant
predictors and revealed compound behavioral-mechanical effects
in Critical-risk cases. The results establish a reproducible synthetic
benchmark and interpretable modeling pipeline, while external
validity remains contingent on future validation with real Dhaka
bus data.

Automated Bug Detection, Root-Cause Theme Discovery, and Solution Recommendation Using Large Language Models

Issue tracking repositories aggregate thousands of natural language bug reports every month, yet screening them into genuine defects, identifying probable root-cause themes, and recommending possible fixes still consumes a disproportionate share of maintenance effort. This paper introduces a comprehensive pipeline that integrates sentence embeddings, dense retrieval, unsupervised clustering, gradient boosted classification, and a retrieval augmented open weight large language model (LLM) to support three interrelated software maintenance tasks: (i) distinguishing between bugs and non bugs, (ii) discovering exploratory root-cause themes, and (iii) recommending solutions in natural language. The system’s performance is assessed using 30,511 cleaned bug reports sourced from four long lived, structurally diverse open source projects in the GitBugs corpus: Apache Cassandra, Apache Hadoop, Mozilla Firefox, and Apache HBase. A Sentence Transformer/FAISS retrieval module reaches an overall nDCG@5 of 0.669, while a LightGBM classifier trained on embeddings, metadata, and structural text flags detects genuine bug reports with 0.8558 test accuracy, an F1-score of 0.9084, and a ROC-AUC of 0.8855; a soft ensemble with a TF-IDF baseline pushes F1 to 0.9100. In an evaluation of 100 held out reports, a retrieval augmented Qwen2.5-1.5B-Instruct model generates root cause hypotheses and fix suggestions. 99 outputs yield parseable bug verdicts, of which 53.5% agree with the resolution derived labels. SHAP explanations expose the metadata and semantic signals on which the classifier depends, and per project confidence intervals show that performance is markedly uneven across repositories. All code, figures, and result artifacts are generated programmatically to support reproducibility.

Early Detection of Traffic Congestion Before Breakdown: A Spatiotemporal Warning Framework for Intelligent Transportation Systems

Proactive traffic management requires more than
identifying congestion after it has formed. This study develops
a multi-horizon early-warning framework for detecting freeway
congestion before confirmed traffic breakdown using real-world
detector data from the Caltrans Performance Measurement
System. The PeMS08 benchmark contains 17,856 five-minute
observations from 170 detectors with traffic flow, occupancy,
and speed measurements. A sensor-adaptive congestion definition
combines speed below 75% of detector-specific free-flow speed,
elevated occupancy, and at least 15 minutes of persistence.
A 30-minute clear period is additionally required before each
onset. Chronological train, validation, and test partitions are
used to prevent temporal leakage. Four classifiers are evaluated
for warning horizons of 5, 10, 15, 20, and 30 minutes. Extra-
Trees is selected using validation precision-recall area under the
curve (PR-AUC). On the held-out test period, it achieves ROCAUC/
PR-AUC values of 0.930/0.825 at 5 minutes and 0.841/0.627
at 10 minutes. Performance decreases with longer lead time,
while useful predictive information remains detectable at 15–30
minutes. Interpretation and ablation analyses show that speed
dynamics, occupancy buildup, and neighboring-detector conditions
provide the dominant warning information. The results
support a 10-minute horizon as a practical compromise between
actionable lead time and predictive reliability for proactive
intelligent transportation systems.

Developing and Validating a Decision Analytics Maturity Model for Human Resource Management Using CFA and Machine Learning

This study proposes and validates the Decision Analytics Maturity Model for HRM (DAMM-HR). The five maturity stages of the model are reactive reporting, diagnostic insight, predictive foresight, prescriptive optimization, and autonomous decisioning. Firstly, it offers a theoretical contribution by establishing decision embeddedness as a key aspect of HR analytics maturity. Second, it does offer an empirical contribution by conducting multi-phase validation via confirmatory factor analysis, ordinal regression, and qualitative text analysis based on data from 217 HR analytics
practitioners. Third, it provides a practical diagnostic tool to help HR leaders assess their organizations’ maturity and identify areas for further development.

Data Leakage and the Limits of Attention: A Significance-Tested Benchmark for Solar Panel Fault Detection

We present a reproducible, leakage-free benchmark evaluated on two complementary PV
inspection datasets: a ground-level RGB surface-fault dataset and a large-scale aerial thermal anomaly dataset. We systematically compare several attention mechanisms integrated with an EfficientNet-B0 backbone under a rigorous evaluation
protocol based on 5-fold cross-validation, multiple random seeds, and statistical significance testing. Across both modalities, no attention module yields a statistically significant improvement over the plain backbone, while a controlled experiment shows that augmenting before splitting inflates test accuracy by approximately ten percentage points. The findings highlight the importance of leakage-free evaluation and statistical validation, and establish a reproducible benchmark for fair comparison of future PV fault detection methods.

Beyond SOH: A Composite Degradation Index with Uncertainty-Aware SOC Estimation for EV Batteries

Conventional battery health indicators such as State of-Health (SOH) capture only capacity fade, while neglecting operational stress, round-trip efficiency losses, and model-derived uncertainty. All of these features accelerate aging in real world electric vehicle (EV) environments. Extending our prior uncertainty-aware State-of-Charge (SOC) estimation work, this paper introduced a novel degradation index named the Unified Degradation Score (UDS), a Ridge-regularized composite metric integrating Depth-of-Discharge (DOD), SOH, C-rate, Round-Trip Efficiency (RTE), and Monte Carlo (MC) Dropout epistemic uncertainty. Validated on the CALCE CS2-36 dataset
over 729 controlled aging cycles, the underlying FT-Transformer achieved MAE 0.599% and RMSE 0.846% during early-life operation. Beyond cycle 250, epistemic uncertainty rose systematically, functioning as an automatic model-derived degradation signal.The UDS achieved a Pearson correlation of r = −0.975 with SOC prediction error, outperforming SOH alone (r = −0.938).A closed-loop BMS charging simulation demonstrated that UDS triggered current throttling preserved 79.0% remaining useful life (RUL) at cycle 729, surpassing conventional SOH-step control (77.2%) and unmitigated fixed-rate operation (64.1%).

Why Fixed-Weight Fusion Fails: A Non-Circular Evaluation of LSTM–DBSCAN–XGBoost for Predictive Maintenance

This paper’s real contribution isn’t the LSTM+DBSCAN+XGBoost architecture itself — those are established methods. It’s a methodological finding, made visible only by a non-circular evaluation protocol:
Fixed-weight fusion of a forecaster, an anomaly detector, and a health classifier looks catastrophic (macro-F1 0.592 vs. 0.913 for the best single branch) but that’s almost entirely a measurement artifact, not a real weakness of fixed weights. Once you fuse at the probability level instead of collapsing everything to a scalar first, the same fixed weights recover to 0.892. 0.298 of the 0.322-point gap comes from throwing away class information (scalarization); only 0.024 comes from the weights being fixed rather than learned (statistically confirmed, p=0.003).

Twenty-Five Years of Air Quality in Bangladesh: Trends, Seasonality, and Spatial Pollution Regimes

This study provides a comprehensive national-scale analysis of air quality in Bangladesh by examining 25 years of hourly data covering over 3.19 million records across 103 cities. It reveals that the apparent flat long-term national trend is an artifact of network expansion diluting severe pollution in hotspots like Dhaka and Narsingdi, which worsen at a rate of 1.84 AQI points per year. Furthermore, the study identifies strong, statistically significant seasonality with peaks in January and lows in July, maps a clear urban-rural gradient showing air quality improvement with distance from the capital, categorizes cities into four distinct pollution regimes via K-means clustering, and establishes that particulate matter PM2.5 and PM10 serve as the primary driver of overall AQI variation.

Advanced Analysis of Amplitude Modulation and Demodulation Techniques for Enhanced Signal Processing in Analog Communication Systems

mplitude Modulation (AM) remains a
fundamental technique in analog
communication systems, essential for
transmitting information over long
distances. This paper presents an
in-depth analysis of AM and its
demodulation processes, focusing on the
impact of varying message and carrier
frequencies on signal fidelity and
modulation efficiency. Using a hands-on
experimental approach, AM signals were
generated and observed under controlled
conditions, followed by a detailed
examination of the demodulation process.
Results indicate that the modulation index
significantly influences the amplitude and
quality of the transmitted signal, while
sideband frequencies play a crucial role in
maintaining signal integrity. The study
further explores challenges encountered
in maintaining optimal signal modulation
and discusses potential enhancements to
conventional AM techniques. These
findings provide valuable insights into
improving analog signal processing, with
implications for enhancing communication
reliability in AM-based systems.

Low-Cost Water Level Indicator Using Logic Gates for Enhanced Water Management

Water scarcity and tank overflow pose significant
challenges in urban Bangladesh, leading to substantial water
wastage. This paper presents a cost-effective water level indicator
designed for domestic water management, utilizing a single IC
74HC04 NOT gate, two BC547 transistors, LEDs, and resistors.
The system leverages water’s conductivity to detect mid and
upper water levels, providing visual indications via LEDs at a
total cost of approximately 300 Taka ( 2.46 USD). Compared to
alternative solutions like ultrasonic sensors ( 1200 Taka) or float
switches ( 950 Taka), the proposed design offers a 3–4 times cost
reduction, making it accessible for low-income households and
educational applications. Implemented on a portable breadboard
setup, the system ensures reliability, with stable LED indications
verified through experimental testing. This work provides a simple, affordable solution tailored to resource-constrained settings,
promoting efficient water management in urban environments.

Implementation of an FM Transmitter Circuit for FM Radio Signal Communication

This paper presents the design and implementation
of a three-stage FM transmitter for short-range audio broadcasting within the 88-108 MHz FM band. The system consists of
three primary stages: pre-amplification, carrier generation, and
power amplification. Initially, the weak audio signal is amplified
by Q1, a NPN transistor, to a level suitable for modulation. In
the next stage, Q2, configured as a Colpitts oscillator, generates
a stable high-frequency carrier signal. This carrier is modulated
by the audio signal in the FM stage, where the frequency of
the carrier is varied in proportion to the amplitude of the audio
input, encoding the audio information.The modulated signal is
then amplified by Q3 in the power amplification stage, ensuring
sufficient strength for short-range transmission. The amplified
FM signal is passed to the antenna, which broadcasts the signal
for reception by FM radios. The system was validated using
MATLAB simulations, which modeled the frequency response,
power spectral density (PSD), and the effects of Additive White
Gaussian Noise (AWGN) and multipath interference on the signal.
The Proteus simulation environment was used to further verify
the circuit’s performance in real-world conditions. The results
demonstrate that the transmitter is capable of handling signal
interference, maintaining frequency stability, and delivering a
robust FM signal for broadcasting.This work provides a novel
approach to FM transmission, integrating theoretical analysis,
simulation models, and practical validation. It addresses common
challenges in wireless communication, such as signal clarity, frequency drift, and interference, making it a valuable contribution
to the field of short-range FM broadcasting and educational
applications

Wideband Antenna Design and Performance Analysis Using CST Simulations

This paper has presented the design and analysis
of a wideband rectangular-shaped antenna using FR-4 as the
substrate, optimized for ultra-wideband (UWB) applications. The
antenna’s performance has been measured using the reflection
coefficient (S11) through 3 GHz to 11 GHz, with CST simulations
which have shown an excellent impedance match, achieving a
minimum S11 of −24.84 dB at 9.44 GHz and −20.95 dB at 4
GHz, indicating good impedance matching at lower frequencies.
Notably, S11(5) has demonstrated near-perfect impedance matching with S11 = −48 dB at 9 GHz, making it suitable for UWB and
advanced wireless systems. When the surface current at 4 GHz
and 9 GHz has been analyzed, it has been found to be evenly
spread out at 5 GHz for broad radiation and concentrated at
9 GHz for higher directivity. These results have demonstrated
how well the antenna has performed over a wide range of
frequencies. This paper has demonstrated the antenna’s superior
impedance matching and wideband performance, applicable in
indoor positioning, short-range communication, radar sensing,
high-speed data transmission, through-wall radar, and medical
or security imaging, thereby underscoring the significance of
simulation-based design in advanced wireless technologies

Adaptive Filtering Techniques for Signal Enhancement in Modern Communication Systems

Signal transmission is central to modern communication systems because reliable, high-fidelity data transfer
depends on it. As demand for higher data rates and stronger performance grows, quantization noise an unavoidable by-product
of analog-to-digital conversion (ADC) has become an obstacle.
Traditional static noise-shaping filters can reduce this noise but
are ill suited to changing signal and channel conditions. This
paper presents an Adaptive Sigma-Delta Modulator (ASDM), a
software configurable scheme that adjusts its noise-shaping filter
in real time to the prevailing signal and channel state. By refining
the modulation process, the ASDM lowers quantization noise
and thereby raises the Signal-to-Noise Ratio (SNR). Simulation
studies show that, under dynamic conditions, the ASDM outperforms conventional Sigma-Delta Modulators (SDMs), delivering
marked gains in dynamic range, stability, and system performance. The design also integrates adaptive filtering including
the Least-Mean-Squares (LMS) algorithm to enable continuous
rate and phase-aware adjustments that improve signal quality
in real time. Beyond performance, the approach offers practical
benefits: higher power efficiency, simpler hardware, and robust
handling of non-stationary signals. These attributes make the
ASDM a strong candidate for high-resolution ADCs, wireless
communication systems, and other applications requiring precise
signal processing. This dynamically responsive strategy points
toward stronger, cost effective, and adaptive solutions for nextgeneration communication technologies.

Archives: Paper Submission