Paper Submission - NPS Australia Submission System

Hybrid GraphSAGE–ANN for Lung Cancer Risk Prediction Using Patient Similarity Graphs

This paper proposes a novel Hybrid GraphSAGE–ANN model that fuses graph-based patient-similarity representations with non-relational feature learning for lung cancer risk prediction. Benchmarked against ten baseline and hybrid GNN architectures on a 5,000-patient dataset under a unified, leakage-safe evaluation protocol, the proposed model achieves the best overall performance (98.80% accuracy, 97.53% F1-score, 99.96% ROC-AUC) and the strongest clustering-agreement scores among all graph-based models, while uniquely offering an interpretable patient-similarity graph for clinical cohort analysis.

RobustEWaste: A Data-Centric Robustness Evaluation of Lightweight E-Waste Detection Under Simulated Smart-Bin Conditions

Electronic waste is the stream of discarded electrical and electronic equipment, and e-waste detection automates its sorting by having a camera recognize items such as phones, batteries, and circuit boards. This paper measures how much of a lightweight detector’s accuracy survives the imaging conditions a real smart bin would impose. Published e-waste detectors report strong accuracy, but almost all are trained and tested on a single dataset under uniform conditions, so their reported figures say little about performance under the poor lighting, occlusion, and compression a deployed camera actually encounters. We present RobustEWaste, a data-centric evaluation framework that harmonizes inconsistent public label vocabularies into an 8-class taxonomy, audits the result for duplicates and annotation defects, applies a reproducible software corruption pipeline covering low brightness, glare, shadow, motion blur, compression, rotation, and occlusion, and reports a Robustness Drop metric alongside standard accuracy. On YOLO11s, smart-bin augmentation with class-balanced oversampling raised clean mAP@0.5 from 0.723 to 0.830 and corrupted mAP@0.5 from 0.296 to 0.356 at no inference cost, running near 90 FPS on a T4 GPU. Robustness Drop nonetheless stayed above 57 percent for both configurations, and the errors were overwhelmingly missed detections rather than class confusions. Clean-set accuracy therefore overstates deployable performance by more than half.

A Multi-Seed Comparative Study of Active Learning Query Strategies for Bangla Product Review Sentiment Analysis

This study provides a controlled multi-seed comparison of four active-learning query strategies for annotation-efficient Bangla product-review sentiment analysis using a fixed human-adjudicated gold test set. It evaluates not only final Macro F1 but also full learning trajectories, AULC, cross-seed stability, and class-wise behavior, showing that more complex acquisition strategies do not necessarily outperform Random sampling and highlighting a persistent Neutral-class bottleneck.

TS-GAT+FC+Adv: Connectivity-Guided Graph Attention for EEG Schizophrenia Detection under Subject-Independent Evaluation

Schizophrenia is a chronic psychiatric disorder diagnosed almost entirely through clinical interview. Early detection matters, since longer untreated psychosis is associated with poorer outcome, yet no objective electrophysiological marker is in routine use. Deep learning on resting-state electroencephalography has been reported to separate patients from controls with very high accuracy, but most such reports do not state how the training and test data were divided. Where epochs from one recording fall on both sides of that division, a network can succeed by recognising the individual rather than the illness. Before such models can be considered clinically, the field needs evidence of how they behave on unseen people. We provide that evidence on a public cohort of twenty-eight participants, half diagnosed with paranoid schizophrenia, recorded over nineteen scalp channels. Four architectures were trained under one common budget and evaluated twice. Under record-level cross-validation, all four exceed ninety-nine percent and are indistinguishable. Under participant-grouped cross-validation of the same trained models, accuracy falls to the mid-to-upper seventies. Our graph attention network, which guides inter-electrode attention by scalp geometry and channel coupling and is trained adversarially against participant identity, attains the best grouped accuracy, the smallest fall between protocols and the most stable estimates. Paired testing nevertheless separates none of the four models, so the protocol effect is the firm finding.

Zero-Knowledge Proof Based Multi-Factor Authentication System for IoT Devices

The key contributions of this work are summarised as
follows:
• A working ZKP-based multi-factor system. We im-
plement the interactive Schnorr protocol end to end and
compose it with server-side face matching and an email
possession factor across five selectable modes, with an
ESP32-CAM as the physical authentication endpoint.
• A credential store with no recoverable secret. The
verifier persists only Y = gx mod p. Section III states
precisely what this protects against and, equally impor-
tantly, what it does not: a short PIN remains subject to
offline enumeration.
• An operating-point finding. Benchmark evaluation
shows impostor similarity never exceeding 0.167 while
the deployed threshold is 0.65, so the threshold can be
lowered substantially without admitting a single false ac-
ceptance. This is a correction to our own design obtained
by measurement rather than a property we designed in.

Region Weighted Conditional GAN for Human Skin-Focused Grayscale Image Colorization

This work addresses a failure mode that generic colorizers share but do not treat: regions whose plausible colors are narrowly constrained. Human skin admits only a small palette, yet in our experiments it was the region most often colored unrealistically or left uncolored. We introduce a region-weighted reconstruction loss that adds a masked L1 term over the skin region alongside the standard full-image L1 and adversarial terms, with the total reconstruction weight held constant so that any gain is attributable to the region term rather than to a shifted loss balance. We also contribute a skin mask derivation from the CIHP parsing annotations and a 60,011-image training set that combines CIHP with skin-free Places 365 images, so the model sees both skin and non-skin scenes. Evaluated within the skin region across two generator architectures, the region term improves PSNR, MAE, and SSIM in both pairs, with the best model reaching PSNR 36.98, MAE 1.22, and SSIM 0.99, and receiving the highest ratings in a blind human review. A secondary finding is that FID and Inception Score mildly favor the unfocused models even as human reviewers prefer the focused ones, indicating that distribution-level metrics do not capture perceptual gains confined to a small region.

Adaptive Paracetamol Exposure Risk Classification with Cost-Aware Information Acquisition

Paracetamol is widely used worldwide, yet substantial inter-individual pharmacokinetic variability can lead to markedly different exposure following similar dosing. This paper presents a computational framework for adaptive, cost-aware classification of elevated paracetamol exposure in resource-limited settings. A virtual population of 10,000 patients incorporating age, sex, body weight, serum creatinine, and creatinine clearance (CrCl) was generated, and 48-hour concentration profiles were simulated using a one-compartment pharmacokinetic model. Toxic Exposure Time (TET) above 20 mg/L was used as the model-defined exposure endpoint. Sensitivity analysis identified clearance as the dominant driver of TET (PRCC = −0.573, p < 0.001). Information ablation showed modest discrimination from freely observable variables (AUC = 0.712), while inclusion of CrCl increased AUC to 0.731. The full model achieved AUC = 0.728. A two-stage adaptive acquisition strategy, using additional CrCl information only for uncertain cases, achieved AUC = 0.735 while reducing testing expenditure by 88% relative to universal testing. However, sensitivity remained low, indicating that the available feature set is insufficient for reliable identification of elevated exposure. The results demonstrate both the potential and limitations of cost-aware adaptive information acquisition for pharmacokinetic risk classification and motivate future validation with additional clinically informative biomarkers.

Techno-Economic Analysis of Hydrogen Storage- Based Solar Power Systems for Off-Grid Industrial

This research proposes a space-efficient solar–hydrogen hybrid energy system for Bangladesh’s industries, addressing limited space and continuous 24-hour power demand. Its novelty lies in the industrial-scale integration and optimization of PV with hydrogen storage and fuel cells under Bangladesh-specific conditions, an area that has received limited attention. The system also compares hydrogen and battery storage using multiple optimization algorithms to achieve reliable and cost-effective operation.

An Explainable Speech-Based Conversational AI System for Sustainable, Automated English-Speaking Assessment and Personalized Feedback for EFL Learners

Automated spoken-language assessment is essential to intelligent educational systems, particularly for scaling sustainable, high-quality feedback in resource-constrained EFL settings. Yet most conversational AI tutors provide dialogue without explainable, learner-specific speaking evaluation. We propose an Explainable Speech-Based Conversational AI Framework that combines Whisper-based automatic speech recognition (ASR), multilevel speech and linguistic feature extraction, and large language model (LLM) feedback in a transparent pipeline. The system transcribes learner speech, measures fluency, lexical, grammatical, and pronunciation features, and maps detected problems to evidence-grounded explanations and personalized recommendations through structured LLM prompting. We evaluate the framework with 60 university-level EFL learners in a six-week mixed-method study: 30 used the AI system and 30 followed conventional speaking practice. Compared with ratings from three certified instructors, AI feedback achieved Cohen’s kappa = 0.74 for weakness identification. The experimental group showed significantly greater gains in fluency (Cohen’s d = 0.86) and grammatical accuracy (d = 0.79). By automating high-quality, evidence-grounded feedback, the framework offers a scalable, sustainability-oriented approach to reducing reliance on limited human-instructor resources.

Socially Responsible Technology: Examining Social Media Addiction and Mental Health among University Students

The expanding adoption of social media offers new opportunities for information sharing and social connection, yet excessive use can adversely affect the psychological well-being of university students, making ethical digital engagement essential for both social sustainability and human capital development. Despite growing global evidence, limited research in Bangladesh has examined how social media addiction relates to depression, anxiety, fear of missing out (FoMO), escapism, loneliness, and self-esteem. This study assessed the prevalence of social media addiction and its association with mental health indicators among university students in Bangladesh. A cross-sectional survey was conducted with 300 students drawn from 15 public, private, and national universities, using validated psychometric scales and multiple linear regression analysis. Results indicated that social media addiction, loneliness, FoMO, and escapism significantly predicted both depression and anxiety, explaining 50.4% and 53.7% of the variance, respectively. The findings highlight the need for responsible digital engagement, balanced lifestyles, counseling, and spiritual practices. University students should adopt conscientious social media habits to mitigate depression and anxiety, support mental well-being, and reinforce human capital development and social sustainability.

Latency-Aware Federated Learning for Network Intrusion Detection under Non-IID Data

Federated Learning (FL) enables privacy-preserving network intrusion detection without sharing raw traffic, yet its deployment on resource-constrained edge environments remains limited by non-IID data, communication overhead, latency, and privacy–utility trade-offs. This paper presents a systems-level evaluation of FL-based intrusion detection on CIC-IDS-2017 under Dirichlet non-IID partitioning (α = 0.5), 10 clients, and 50% client participation. We systematically compare FedAvg, FedProx, SCAFFOLD, and FedNova with a momentum-stabilized FedAvgM using a LayerNorm/SiLU client model, evaluating detection performance, convergence, latency, communication cost, and energy consumption. FedAvgM achieves 98.44% accuracy and
98.16% Macro-F1, reaching its best performance by round 58, outperforming the evaluated federated baselines. A componentwise ablation confirms the complementary benefits of LayerNorm and server-side momentum, while DP-SGD quantifies the resulting privacy–utility trade-off, achieving ϵ = 5.7343 at δ = 10^−5. The results demonstrate that momentum-based aggregation can improve both convergence and detection performance under partial participation and non-IID data, while highlighting the
practical cost of formal differential privacy in FL-based intrusion detection.

Recurrent Graph Reinforcement Learning Framework for Task Offloading in Internet of Vehicles

The rapid growth of the Internet of Vehicles (IoV) has increased the demand for low-latency, real-time, and priority-aware applications recently. Current Mobile Edge Computing (MEC) enabled vehicular architecture often fails to address these challenges effectively, which leads to lower quality of service. In this paper, we propose a joint mobility-aware task offloading and priority-based resource allocation framework for MEC-enabled IoV. We formulate a priority-aware multiobjective optimization problem to minimize latency and energy consumption of the offloaded task, which is a mixed-integer nonlinear problem (MINLP) and NP-Hard to solve. In this regard, we develop a Recurrent Graph Reinforcement Learning (RGRL) framework integrated with a Graph Neural Network (GNN). The GNN captures the spatio-temporal relationships among vehicles, Roadside Units (RSU), and edge servers, enabling intelligent and adaptive decision-making in dynamic environments. Simulation results demonstrate that the proposed RGRL-GNN approach outperforms state-of-the-art works in terms of latency reduction, energy efficiency, and successful task completion rate.

Revisiting Behavioral Android Malware Classification: A Leakage-Controlled Benchmark of Tree Ensemble Methods and Hybrid Attention Networks

Presents a leakage-controlled benchmark comparing strong tree ensembles with a hybrid attention network on behavioral Android malware data. The study shows that well-tuned tree ensembles outperform the more complex HAT-Net while maintaining strong accuracy and efficiency, providing a rigorous and reusable evaluation methodology for tabular malware detection

Reliability and Conflict-Aware Closed-Loop Multimodal Edge Framework for Landslide Early Warning: A Simulation-Based Algorithmic Evaluation

This paper presents a reliability- and conflict-aware multimodal edge fusion framework for landslide early warning under sensor degradation. The framework treats modality reliability and cross-modal disagreement as separate factors and adjusts sensor contributions using availability, range validity, temporal stability, and drift. A frozen chronological evaluation protocol is developed, with warning thresholds selected under a validation-stage false-alarm constraint, using a real landslide inventory from three regions of Bangladesh. On the locked test set, the proposed method achieves 66.7% episode recall while producing fewer false alarms than rainfall-only and conventional fusion. Leave-one-episode-out evaluation across 16 reconstructed episodes further indicates higher recall and lower false-alarm rates. The study provides a reproducible framework for assessing robust multimodal edge warning under controlled sensor degradation, while distinguishing algorithmic robustness from operational forecasting performance.

HP-MMFA: Mobility-Aware Min-Max Resource Allocation for Vehicular Cloud Computing

With the rapid development of communication technologies in Autonomous Vehicular Clouds (AVC), modern vehicles possess increasing amounts of idle storage and computing resources that are often unused and can be shared with nearby vehicles to support different services. In vehicular cloud computing, many services are elastic, which means they can continue operating with partial resources, although with reduced quality or slower execution. Examples include video streaming, data upload, sensing data transfer, and non-urgent computation tasks. For such services, it is more practical to guarantee a minimum acceptable resource level and then provide additional bandwidth when more resources are available.

For this reason, Min-Max resource allocation is suitable for elastic vehicular services. Each admitted Client Vehicle (CV) first receives its minimum required bandwidth so that the service can continue operating. The remaining bandwidth is then distributed progressively up to the maximum requested bandwidth of each CV. This permits more vehicles to remain in service under limited resources. Therefore, elastic services benefit from the Min-Max method by improving service continuity, resource utilization, and fairness.

The main contributions are:
A mobility-feasibility model considering RSU residence time and CV–SV link duration.
HP-MMFA, which prioritizes feasible handoff tasks with smaller mobility slack before Min-Max allocation.
Comparison of B-MMFA, LP-MMFA, and HP-MMFA using 100 simulation runs.

An Efficient Protocol for Private Face Recognition Using Post-Quantum Secure Encryption

We propose an efficient private face recognition protocol using CKKS fully homomorphic encryption that addresses SIMD slot waste in per-embedding CipherFace baselines. By row-packing multiple face embeddings into contiguous blocks within a single ciphertext and applying a mask-free block-isolated hierarchical reduction, our method enables parallel distance computation over packed embeddings. Experiments on the LFW dataset with FaceNet, FaceNet512, and VGG-Face embeddings demonstrate significant speedups over the CipherFace baseline, with Hamming distance on binarized embeddings offering additional efficiency while maintaining comparable accuracy at d=512.

Influence of Color Space on GAN-Based Image Colorization with a Grayscale to Color Network

This work isolates a variable that the colorization literature has consistently left uncontrolled. Prior GAN-based colorizers each pick a color space, but the choice is rarely justified, and the few papers that do compare spaces vary the architecture at the same time, so the effect cannot be attributed. We hold the cGAN generator, discriminator, loss function, optimizer, and training schedule fixed and vary only the color space across RGB, Lab, HSV, YUV, and YCbCr. The result is a clear and reproducible trade-off: YCbCr yields colors closest to the ground truth (PSNR 21.29, MAE 16.42) while Lab yields the structurally smoothest output (SSIM 0.87), and a blind review with 78 participants confirms both as the preferred spaces. We then apply this finding to select a generic colorizer, showing that a ResNet-50 encoder generator trained in Lab under a generator-ahead schedule reaches PSNR 22.38, SSIM 0.90, and FID 45.88 on DIV2K, competitive with published methods. The contribution is therefore a controlled empirical basis for a design decision that is currently made by convention.

ResQPlus: A Weapon Detection and Emergency Alert Platform for Real-Time Public Safety Using YOLO11n

This work presents ResQPlus, a unified four-layer IoT reference architecture (perception, edge processing, network/communication, application services) that integrates real-time weapon detection with automatic evidence capture, GPS-based location sharing, and multi-stakeholder emergency alerting; a combination rarely addressed as a single coordinated pipeline in existing literature, which tends to treat detection accuracy and alert routing as separate problems. The perception layer was validated using a lightweight YOLO11n model, achieving 89.66% precision and 80.30% recall at 8–45 ms inference latency. More significantly, the complete detection-to-alert pipeline from camera-based threat identification through cloud routing to simultaneous police dispatch and citizen notification which completes in under two seconds end-to-end, demonstrating practical viability for time-critical, resource-constrained surveillance deployments across homes, schools, vehicles, and workplaces.

Domain-Adaptive Fusion of Wavelet-Entropy and wPLI Graph Attention for Cross-Dataset EEG Seizure Detection

Cross-dataset EEG seizure detection remains challenging because models trained on one dataset often degrade under differences in patient populations, acquisition settings, and background EEG distributions. This paper proposes a domain-adaptive dual-stream attention-fusion framework for bidirectional seizure detection between the CHB-MIT and Siena Scalp EEG datasets. The local stream combines wavelet-entropy features with channel-attention CNN learning, while the global stream models inter-channel functional connectivity using weighted Phase Lag Index graphs and an edge-aware graph attention network. Their embeddings are integrated through sample-adaptive attention fusion, and a domain-adversarial network with a gradient reversal layer uses unlabeled target-adaptation data to reduce source-target mismatch. Under patient-disjoint evaluation, the framework achieves 92.84% accuracy, 90.75% F1-score, and 92.96% AUC for CHB-MIT→Siena, and 91.34%, 89.42%, and 91.18% for Siena→CHB-MIT. Domain adaptation improves AUC by 2.60 and 2.85 percentage points, while post-hoc domain-probe accuracy decreases from 94.1% to 55.3%, indicating reduced dataset separability while preserving seizure-class structure. These findings support complementary local-global representation learning and adversarial alignment for robust cross-dataset EEG seizure detection.

Interpretable Brain Tumor MRI Classification: Comparing GLCM-Based Classical Models and CNNs

This work presents a direct, controlled comparison between handcrafted GLCM-based texture features (with Logistic Regression and Random Forest) and an end-to-end CNN for four-class brain tumor MRI classification, using identical data splits and preprocessing. The CNN achieves 92.00% accuracy and a 91.54% macro-F1 score, substantially outperforming both classical baselines. Grad-CAM is further applied to interpret CNN predictions, with specific focus on the glioma–meningioma confusion pair, offering qualitative insight into the model’s spatial decision-making.

Serial RAG Engine for Memory-Efficient Offline Math Learning

Mathematics learning still poses challenges to most
students owing to a lack of personalised guidance, fewer learning materials, and inadequate access to intelligent tutoring systems,especially in poor-resource environments. In this paper, we propose MathAssist, an offline AI-assisted mathematics learning framework that offers context-aware and stepwise mathematical support based on multimodal input, RAG, and lightweight Large Language Models (LLMs). The framework accepts text, image,and voice input and translates them to a common representation of a query. A local knowledge base is employed for semantic retrieval, whereas a serial RAG architecture facilitates interaction between retrieval and reasoning processes done via LLMs while avoiding unnecessary concurrent resource consumption. In order to choose a suitable model for the framework, we evaluated nine lightweight open-source LLMs using the MathQuest benchmark, which contains 42 mathematical questions divided among Easy,Medium, and Hard classes. The assessment took into consideration accuracy, efficiency, compatibility, response time, RAM usage, GPU usage, and token usage. WizardMath ranked first with an accuracy of 91.06%, and therefore was chosen as the main model, whereas Phi-3 had the second-highest ranking.Moreover, case-study experiments have been conducted on 3 quadratic equations via a Python and Streamlit-powered testing platform. Live resource monitoring revealed maximum RAM utilisation of 13.02 GB with 61% GPU utilisation for WizardMath and 12.27 GB with 63% GPU utilisation for Phi-3 at the inference stage. The evaluation was conducted in a Python and Streamlit-based environment, and the observed resource consumption indicates that further optimization is required before deployment on low-memory mobile devices.The main contributions of this research include:
• Benchmarking of nine lightweight open-source LLMs
using the MathQuest dataset and selection of WizardMath
as the best-performing model
• Development of a Serial RAG method that combines
semantic retrieval, local knowledge indexing, and LLM
inference.
• Resource consumption monitoring using a Python-based
evaluation environment in the case study of WizardMath
and Phi-3. and evaluation of LLMs using a Python and
Streamlit-based testing environment which lead to depict the performance evaluation of the LLMs.

BanglaSentiEmo: Cross-Task Affective Structure from a Shared BanglaBERT Encoder for Joint Sentiment-Emotion Classification in Bengali

Abstract—Sentiment analysis and emotion recognition are predominantly treated as independent tasks in Bengali NLP, despite both being grounded in the same underlying affective signal. Existing multi-task learning (MTL) studies either pair sentiment with structurally unrelated tasks, or rely on mul-tilingual encoders trained on two disjoint corpora without a genuinely shared encoder. In this work, we propose a shared-encoder BanglaBERT framework for joint sentiment (5-class) and emotion (7-class) classification of Bengali social media text. We first quantify the sentiment-emotion dependence on the full dataset (χ2 = 11090.04, df = 24, p < 0.001; bias-corrected Cramer’s V = 0.684; normalized mutual information = 0.556; information gain = 1.373 bits), confirming a strong association that motivates joint modeling. Under a matched protocol repeated over five random seeds, the shared encoder shows no statistically significant change in sentiment macro-F1 (STL 65.25 ±1.55 vs. MTL 65.39 ±1.45; paired t-test p = 0.893) and a statistically significant regression in emotion macro-F1 (STL 69.21 ±0.87 vs. MTL 67.39 ±1.02; p = 0.023), so raw per-task F1 does
not reliably improve under the joint objective. Despite this, error analysis on a representative run shows correlated task performance: joint (both-correct) accuracy of 55.91% exceeds the 47.13% expected under independence (+18.6% relative), with a Pearson correlation of 0.4084 (p= 3.334 ×10−25) between per-post sentiment and emotion-correctness, indicating the shared representation captures genuine cross-task affective structure even where this does not consistently translate into higher macro-F1.Index Terms—Multi-task learning, BanglaBERT, sentiment analysis, emotion recognition, Bengali natural language processing, shared encoder, low-resource languages.

Efficient Serial RAG Engine for Offline Math Education

The learning of mathematics is still difficult for
most learners because of the lack of personalised instruction and insufficient use of intelligent tutoring systems.. This paper presents MathAssist, an offline AI-assisted framework that provides context-aware, stepwise mathematical support using text-based queries, Retrieval-Augmented Generation (RAG), and lightweight Large Language Models (LLMs). While the architecture is designed to support future multimodal extensions(e.g. image-based OCR and voice input), the current prototype processes text-based mathematical problems. A local knowledge base supports semantic retrieval, while a memory-aware serial execution strategy ensures that retrieval and LLM inference are performed sequentially, avoiding unnecessary concurrent resource usage. Nine lightweight open-source LLMs were evaluated on the MathQuest benchmark (42 problems across Easy, Medium, and Hard levels). WizardMath achieved the highest accuracy (91.06%) and was selected as the primary model; Phi-3 ranked second. Resource consumption was monitored in a Python VS Code and Streamlit test-based environment. While non-quantised models consume ∼12-13 GB RAM during inference, we note that 4-bit quantisation reduces their storage sizes to 1.8 to 4.2 GB, suggesting a plausible path toward lower-memory mobile deployment with further optimisation. The main contributions of this work are:
• An evaluation of nine lightweight LLMs on the Math-
Quest dataset, with WizardMath achieving the highest
accuracy of 91.06%.
• A serial execution strategy that keeps retrieval and LLM
inference separate and explicitly loads and unloads the
model to limit peak memory usage.
• An analysis of resource consumption that reports the
baseline (non-quantised) memory footprint and examines
the potential of 4-bit quantization for reducing the resource requirements of mobile deployment

Automated Difficulty and Algorithmic Tag Prediction for Competitive Programming Problems Using Natural Language Processing

This research proposes a lightweight NLP-based framework for automatically predicting the difficulty rating and algorithmic tags of competitive programming problems. The system combines TF-IDF text features with Random Forest Regression for difficulty prediction and One-vs-Rest Logistic Regression for multi-label tag prediction. The models are evaluated on unseen Codeforces problems, showing that traditional NLP and machine learning techniques can provide useful support for problem analysis, practice selection, and educational recommendation systems.

A Three-Layer Behavioral Authentication Architecture for Resource-Constrained Embedded Devices: Design and Simulation-Based Feasibility Study

This paper presents a low-cost, fully offline three-layer behavioral authentication architecture for resource-constrained embedded devices. The system integrates spoken-digit MFCC-DTW voice verification, ultrasonic hand-withdrawal sensing, and Force–Hold–Gap mechanical interaction on an ATmega328P-class microcontroller. Sequential screening with early rejection and weighted score fusion is evaluated through a held-out Monte Carlo simulation with 50 synthetic users. The study demonstrates architectural feasibility under strict cost, memory, and offline-operation constraints while explicitly identifying the need for physical, human-subject, and adversarial validation.

Attention-Augmented Deep Learning and Decision-Threshold Tuning for Primary Bone Tumor Classification in Radiographs

Primary bone tumors pose a significant oncological
problem wherein timely radiographic diagnosis is crucial. Auto-
mated learning algorithms usually encounter challenges of class
imbalance and overfitting because of small sample size of the
medical datasets. In this paper, we introduce a unifying approach
which uses an attention-based CNN (EfficientNetB4 with CBAM),
focal loss with batch sampling for handling class imbalance and
validation set-based threshold tuning for overfitting mitigation.
We evaluate the proposed approach on the BTXRD dataset and
conduct an investigation of methods which prevent overfitting
such as data augmentation, dropout, weight decay, learning
rate scheduling, early stopping, and label smoothing across five
neural networks (DenseNet121, EfficientNetB3, EfficientNetB4,
MobileNetV3-Large, Swin Transformer-Base) and a soft voting
ensemble. It turns out that default decision threshold (0.50)
produces low precision of the tumor class due to the high test set
imbalance (∼10:1). However, using validation set-based threshold
tuning (θ∗ = 0.8494), it is possible to increase precision from
0.43 to 0.85 while retaining 0.76 recall which results in 96.62%
accuracy, weighted F1-score of 0.9654, specificity of 0.9863, and
AUC-ROC of 0.9760.

Hybrid Deep Learning Framework for Bone Tumor Classification from X-Ray Images Using EfficientNet, DenseNet, and Ensemble Voting

Bone tumor detection using X-ray images is still an
important clinical problem because of the highly heterogeneous
nature of bone lesions in X-rays. In this paper, we propose a
hybrid deep learning model for binary bone tumor classification
using X-ray images. The bone tumor positive X-ray images and
normal bone X-ray images are used as input data for our
task. Three convolutional neural network models are trained
independently: EfficientNetB4, EfficientNetB3, and DenseNet121
using overfitting and without overfitting avoidance strategy. We
propose two hybrid approaches: a feature-level fusion approach
based on EfficientNetB4 and DenseNet121 using Squeeze-and-
Excitation (SE) channel attention module and a hybrid approach
based on EfficientNetB4 with the use of SVM classifiers with
RBF, linear, and polynomial kernel functions. We apply three
ensemble voting schemes, including soft, hard, and weighted
voting methods, for our three backbone networks. For all ten
approaches considered, the best results are obtained by soft and
hard voting ensembles, with 99.19% test accuracy and AUC-ROC
of 0.9996. Visualization with Grad-CAM technique confirms that
the model highlights the regions of interest in the image.

From Messages to Malicious Links: Joint Spam and Phishing Detection via Cross-Attentive Multimodal Learning

* A multilingual Bangla-English-Hindi spam-text corpus containing 11,260 messages and a merged phishing-URL corpus containing 1,248,283 URLs were assembled from six public sources. Five multilingual transformer models and seven classical and deep learning URL classifiers were benchmarked on these datasets as single-task baselines.

* A balanced hybrid dataset of 50,000 message-URL pairs was constructed, covering all four possible spam/phishing label combinations. Five hybrid architectures were then trained by combining a text encoder (MuRIL, mBERT, or DistilBERT) with a character-level URL encoder (GRU, CNN, or BiLSTM) using the same training protocol.

* A hybrid model combining MuRIL and a character-level GRU through multi-head cross-attention was proposed. In this architecture, the text representation directly queries the URL sequence. The proposed model achieved the best overall joint spam-text and phishing-URL performance among the five hybrid configurations evaluated.

* The proposed model was compared with recent spam and phishing detection studies to assess the benefits and limitations of joint text-URL fusion. The comparison showed that hybrid fusion improved performance in several cases, while single-task models remained better in some specialized tasks, including one case where the hybrid URL detection head performed worse than the best standalone URL classifier.

ECG Heartbeat Classification for Cardiac Disease Detection using 1D-CNN with Explainable AI

This paper presents a lightweight (814,215 parameters) 1D-CNN for five-class ECG heartbeat classification on MIT-BIH Arrhythmia dataset with 99% test accuracy and macro-F1 of 0.9195. SMOTE tackles extreme class imbalance without polluting test data, while gradient-based Saliency Maps and handcrafted clinical features (R-peak, QRS duration) offer Explainable AI interpretability a combination not provided by similar previous work on this dataset.

An Intelligent Railway Level Crossing System with Multi-Zone Train Monitoring and Predictive Safety

The research proposes an intelligent railroad level crossing system based on Multi-zone sensors which integrates sensor self-testing, train speed and ETA(Estimated Time of Arrival) estimation, sensor sequence validation, automatic gate control and dual-condition safe exit verification in one framework. The system is meant to improve the safety and reliability of railroad crossings by reducing sensor-related faults and preventing unnecessary gate waiting.

NOTEBRIDGE: A Framework for Real-Time Multimodal Lecture Note-Taking with Evolving Knowledge State

NOTEBRIDGE is a 9-layer real-time multimodal lecture note-taking framework that addresses critical gaps in existing AI educational tools. Unlike systems producing flat transcripts, NOTEBRIDGE maintains an evolving topic-aware knowledge state with hierarchical tracking (supporting suspension, resumption, and merge), grounds claims within authorized curriculum resources via FAISS retrieval, and verifies assertions through a non-mutating verification protocol (F1=0.76). Evaluated on 80 lectures across 5 domains in English and Bangla, it achieves ROUGE-L F1=0.82 and Topic Boundary F1=0.85. Ablation study demonstrates incremental improvement from ASR-only (0.38) to full system (0.81). The framework integrates real-time streaming, privacy-preserving ephemeral processing, and bilingual support for educational note-taking.

A Critical Technical Analysis of Adaptive Privacy-budget Allocation in Federated Medical Imaging

This paper presents a critical technical analysis of adaptive privacy-budget allocation in federated medical imaging, supported by a structured
review of approaches published between 2015 and 2025. Twenty-two studies are examined, including seven addressing non-uniform allocation or related alternatives, focusing on allocation basis, protection granularity, conserved privacy quantity, and accounting compatibility. A reported allocation rule is analyzed using simulated radient-inversion similarity. Under the reported formulation, normalization by client size reduces the difference between the largest and smallest budgets, with the budget ratio approaching 1 and reaching approximately 1.0004 at a client size of 2172, rather than the intended eβ. A rank-normalized variant shows that preserving the sum of per-sample budgets does not preserve Rényi privacy cost, producing excess costs of
2.1%, 8.2%, and 31.3% for β = 0.5, 1, and 2, respectively. Using a reference accounting configuration constructed from available information, the least-protected record yields ϵ = 3.24 for a nominal ϵ = 2 at β = 1. The review identifies limitations in attack-derived sensitivity validation, privacy evaluation, and reporting of parameters required for reproducibility. Overall, the findings demonstrate that adaptive privacy allocation must be evaluated jointly with the conserved privacy quantity and the accountant used to establish privacy guarantees.

SAFER-RAG: Explainable and Calibrated Hallucination Detection with Selective Prediction under Distribution Shift

This work proposes SAFER-RAG, a lightweight and interpretable hallucination detection framework that integrates lexical/factual, semantic, and NLI-based evidence—claim features with confidence calibration, selective prediction, and TreeSHAP explainability. Beyond in-domain detection, the study evaluates reliability under cross-task, cross-generator, source-domain, and external benchmark shifts. The results show that discrimination can remain strong even when calibration degrades, highlighting the need to assess confidence reliability separately. On RAGTruth, Platt scaling reduced ECE from 0.0920 to 0.0583, while abstaining on the least-confident 20% of predictions reduced error risk by 23.6%.

QUBO-Inspired Simulated Annealing for Job-Shop Scheduling: A Benchmark Comparison with Classical Dispatching Heuristics

The paper’s significant contribution is the development of a hybrid quantum-inspired scheduling approach that combines simulated annealing, QUBO-based search, and a deterministic decoder to improve Job-Shop Scheduling solutions. Compared with traditional dispatching heuristics, the method reduced the average gap from 19.32% to 10.27% and achieved better results in 6 of 7 benchmark problems**, while maintaining reasonable computational cost.

MediNetLite: A Resource Efficient Ultra Lightweight CNN for Medicinal Leaf Recognition in Bangladeshi Weed-Infested Area

The main contributions of this work are:

Field-Oriented 34-Class Dataset: To address the key challenge of real-world, cluttered environment medicinal plant recognition, we have introduced a newly compiled dataset of 17050 images captured in the field, comprising 34 medicinal plant species commonly found in weed-infested agricultural areas of Bangladesh.

Custom CNN- MediNet-XG is proposed, which is with 342 KB model size, 98.82% accuracy and 98.81% F1-score which is significantly better than VGG16 (56.20 MB), DenseNet121 (26.98 MB), MobileNetV2 (8.64 MB), and SqueezeNet (889 KB) and is 79×–164× smaller.

Task-Specific Design for Edge Deployment: The architecture utilizes inverted residual blocks with depthwise separable convolutions and lightweight channel attention, achieving both high accuracy and practical deployment on resource-limited devices such as smartphones and edge devices.

Interpretable Multimodal Framework: Integrate Grad-CAM and t-SNE for spatial and feature-space explainability and retrieval-based query module for providing grounded botanical knowledge (uses, safety, dosage) without hallucination for all queries.

Proof-of-concept of Lightweight Superiority: This work demonstrates that in low resource settings, custom ultra-compact, task-specific architectures can outperform generic pretrained models, both in terms of accuracy and efficiency, for fine-grained recognition in agriculture.

Robustness-Aware Fire & Smoke Detection: An XAI-Guided Diagnosis-to-Mitigation Framework

This paper propose an XAI-guided Diagnosis-to-Mitigation (D2M) framework for fire/smoke detection: corruption-specific failures are first diagnosed via Grad-CAM, targeted augmentation is applied only where evidence supports it, and the resulting gains are re-verified through an unbiased XAI attention check rather than accuracy alone. This reveals an asymmetry invisible to standard metrics — Gaussian Noise mitigation is causally verified (+13.78 pp ground-truth-relevant attention), while a similar mAP gain under Low-Light shows no corresponding attention improvement (−1.10 pp) — demonstrating that accuracy alone cannot confirm genuinely improved model reasoning.

A Large-Scale Performance Evaluation of ML-Based and Deep-Neural Models in Evil Twin Attack Identification Using a Hybrid IEEE 802.11 Dataset

The key contribution of this research is a large-scale benchmark of 80 machine learning and deep learning models for Evil Twin attack detection using a hybrid IEEE 802.11 dataset of over 3.8 million samples. It further strengthens the evaluation through independent unseen and public datasets and considers not only detection performance but also generalization, computational efficiency, and deployment suitability, helping identify models that are practical for real-world wireless intrusion detection.

Explanation Faithfulness under Intra-Language Distribution Shift in Low-Resource Bangla Sentiment Classification

1. First evaluation of explanation faithfulness under natural intra-language distribution shift in a low-resource South Asian language (Bangla).

2. All attribution methods show significant raw comprehensiveness drop (32.5–38.7%, p GradientSHAP > Attention) confirms Attention as least robust, with its margin over random shrinking >4× faster than Integrated Gradients.

Longitudinal Alzheimer’s Disease Progression Analysis via a Hybrid Feature-Selected Stacking Ensemble Framework with Explainable AI

This study introduces a leak-free, highly interpretable machine learning framework for tracking Alzheimer’s disease progression using longitudinal MRI data from the OASIS-2 cohort ($N = 373$ visits). The primary contribution lies in formulating dynamic rate-of-change trajectory features ($\Delta\text{MMSE}$, $\Delta\text{nWBV}$) to capture temporal neurodegenerative changes, while embedding a 3-stage hybrid feature selection pipeline and Borderline-SMOTE oversampling strictly within a subject-aware 5-Fold Stratified GroupKFold cross-validation scheme to eliminate intra-subject data leakage. Evaluated across heterogeneous tree-based base estimators and a meta-learned stacking ensemble, the framework achieves high diagnostic performance—peaking at 80.92% precision for XGBoost, 76.68% accuracy for Random Forest, and 0.838 ROC-AUC for Extra Trees—complemented by dual-level XAI (SHAP and LIME) to validate predictions against clinically established neuroimaging biomarkers.

Machine Learning Optimized Microstrip Patch Antenna Array for Application in Pancreatic Cancer Detection

Wide & Deep FAHN: A Fuzzy Attention Hybrid Network for E-Commerce Purchase Prediction with Sequence and Statistical Feature Fusion

1) Dual-Branch Architecture: To develop a Wide and
Deep framework that jointly learns from raw sequential
interactions and engineered statistical session features.
2) Attention and Fuzzy Inference: To integrate attentionbased interaction weighting with Gaussian fuzzy inference for identifying salient behaviors and modeling
uncertain purchase intent.
3) Leakage Prevention and Class Balancing: To establish
a rigorous leakage-safe preprocessing and evaluation
pipeline using undersampling-based class balancing.
4) Benchmarking and Validation: To evaluate the
proposed framework against strong gradient-boosting
baselines using statistical significance testing and
component-wise ablation experiments

Designing a Sustainable Gamified English-Speaking Ecosystem: A Human-Centered Framework for Resource-Constrained Universities

This paper proposes S-GAME, a human-centered and sustainable gamified framework for developing university students’ English-speaking skills in resource-constrained contexts. It contributes by aligning game mechanics with measurable speaking outcomes, embedding accessibility, privacy, teacher control, and low-connectivity support into the system design, and presenting a multidimensional evaluation plan covering learning, engagement, equity, operational cost, and environmental impact.

Understanding How Generative AI Can Encourage Self-Directed Learning: The Roles of Perceived Self-Efficacy and Learning Motivation

he findings of this study are especially important for designing next-generation learning policies and practices integrating GAI tools. As for the theoretical contribution, this study combines human behavioral theories with adult learning theories and extends technology adoption models by exploring the consequences of such adoptions. Practically speaking, educators can find this research especially helpful for understanding how GAI affects students’ learning behaviors and encourages autonomous learning. Tertiary education institutions and EdTech platforms can incorporate the insights of this study in their AI policies and curriculum developments.

An Explainable Deep Learning Architecture for Autism Spectrum Disorder Diagnosis from Functional MRI

First, we integrate ASD classification with SHAP-based connectivity interpretation. Second, we implement a leakage-controlled feature selection pipeline. Third, we map predictive features to candidate anatomical markers using the Harvard–Oxford atlas.

Deep Feature Fusion of CNN and Vision Transformer for Automated Osteoporosis Detection from Knee X-Ray Images

The overall contribution of this research work is as follows:
• Identification of gaps in existing method for knee osteoporosis detection, including limited dataset size, low
images quality, and insufficient feature extraction in conventional (CNN) approaches.
• Application of sophisticated data augmentation methods
to enhance model robustness and generalization on constrained datasets.
• Develop an ensemble based deep learning model that
can do automated diagnostic system using DenseNet169
with a Vision Transformer–inspired attention mechanism,
enabling the model to focus on clinically relevant regions
of knee X-rays.

Sustainable Technology and Energy Justice: Infrastructure and Citizenship in Lavanya Lakshminarayan’s the Ten Percent Thief

Technologically advanced and environmentally resilient urban systems may improve efficiency and material security without necessarily producing socially just outcomes. This paper examines Lavanya Lakshminarayan’s The Ten Percent Thief through the distributional, recognition, and procedural dimensions of energy justice, supported by an Energy Humanities approach to technological infrastructures. Using theoretically informed close reading of the 2023 Solaris edition, the study analyzes spatial division, energy-linked services, climate protection, merit-based citizenship, algorithmic governance, surveillance, and resistant technological practices in Apex City. The analysis finds that infrastructure functions as a civic sorting mechanism: unequal access differentiates material security; Bell Corp’s merit system converts these inequalities into judgments of productivity, dignity, and civic worth; and centralized control of metrics, networks, and decision-making limits political agency. Analog counter-infrastructures, meanwhile, demonstrate technological agency through appropriation, repair, and recombination rather than technological rejection. The novel consequently critiques an exclusionary form of technological sustainability in which technical resilience is separated from equitable access, recognition, and participation. The study shows how speculative fiction can contribute a humanities perspective to debates on socially just sustainable technology.

Conflict Aware Trust-Based Aggregation for Poisoning Resistant Federated Medical Image Classification

Introducing a lightweight conflict-aware trust-based aggregation method which integrates cosine similarity, norm clipping, historical trust memory and softmax-based weighting to mitigate the impact of conflicting or malicious client updates during federated aggregation. Evaluating proposed framework on the PathMNIST dataset with 30\% of persistent malicious clients both in IID and Dirichlet-based Non-IID federated environments to mimic heterogeneous medical learning environments. Comparing proposed aggregation method with FedAvg, Trimmed Mean and Coordinate Median.

Hybrid Inverter-Based Rooftop Solar PV System for Inter-Building Power Sharing: A Techno-Economic Case Study of Green University

The current studies related to rooftop solar PV (SPV) systems on university campuses in Bangladesh are mostly conducted based on a single building, leaving the excess demand of nearby buildings unmet and energy surplus exported to the grid. The present paper proposes a new set up of a 133 JinkoSolar JKM590N-72HL4-BDV bifacial solar array mounted on the rooftop of G-Block at Green University of Bangladesh (GUB) to supply power to both G-Block and adjacent H-Block using a single Huawei SUN2000-75KTL-C1 hybrid inverter with dual-feeder distribution. Based on site-specific Meteonorm weather data, a simulation with PVsyst v7.4.7 is performed, which yields an annual energy yield of 107,102 kWh (specific yield 1,365 kWh/kWp/year, performance ratio 84.2%). The system costs USD 26,787 (USD 0.34/Wp) to install, and has a simple payback period of 4.0 years, and a levelized cost of energy of USD 0.0318/kWh under NPBS-2 tariff of BDT 9.05/kWh, saving around BDT 969,273 (≈USD 7,880) per annum. The system has a net carbon footprint mitigation of 1488.1 tCO2 over a lifetime of 30 years. The results show that a single building’s roof-top PV power can be shared between two neighbouring buildings by a single hybrid inverter which provide favourable technical, financial and environmental outcomes in a low cost and scalable way towards zero-grid building operation.

Synergistic Design of Renewable Energy and Waste Management through Urban Drainage Systems

This paper proposes a dual-pathway urban-drainage system, which includes waste interception, micro-hydropower generation and an AI lighting application, as all of these features were validated using a prototype installation. This concept is scaled up to field scale by a physics based scale-up model based on the actual Dhaka drainage data to provide a renewable energy pathway that is affordable and requires minimal maintenance for waste-constrained, fast urbanizing cities.

Cross-Platform Social Media Analytics for Sustainable Climate Discourse: A Multimodal Analysis of Frame Propagation Across Facebook, YouTube, and Instagram

This research establishes a multimodal analytics framework, while considering platform functionality, to understand climate-related messaging in environmental campaigns on Facebook, YouTube, and Instagram. The research utilizes systematic qualitative content analysis to analyze 60 artifacts from environmental campaigns, including those of the UN Climate Change Secretariat, WWF, and Greenpeace International, collected during 2024–2025. Entman’s framing theory and the multimodal discourse analysis framework proposed by Kress and van Leeuwen are used to model cross-modal relations. The results indicate that the platforms differ in their communicative functions. For example, YouTube enables time-based audiovisual elements to frame disasters and science; Instagram frames campaigns through aspirational images and social media influencers in a hopeful manner; and Facebook facilitates infographics and polyvocal framing of campaigns.

Interpretable Student Academic Performance Analysis Using Hybrid Feature Selection and Explainable Artificial Intelligence

We used hybride feature selection method. There were three feature selection methods: Mutual Information, RFE, and Random Forest. These methods helped us reduce 40 features to 15 important features. Ridge Regression gave the best result with an R² of 0.7353. We also used SHAP to understand which factors most affect student performance.

A Lightweight Plasma Cell Detection Framework with Knowledge Distillation Based Feature Refinement for Multiple Myeloma Diagnosis

This research proposes a lightweight YOLOv8n-based framework enhanced with P3 feature refinement to improve detection performance while maintaining model efficiency. Feature-level knowledge distillation is further investigated to assess the effectiveness of transferring information from a larger teacher model to the lightweight detector. The study also incorporates error analysis and explainable AI to provide deeper insight into model performance and limitations.

From Prohibition to Guided Adoption: A Socio-Technical Framework for Human-Centered Generative AI Governance in Higher Education

This paper identifies a cross-institutional shift from blanket prohibition of generative AI toward guided adoption in higher education. Its main contribution is the proposed Human-Centered GenAI Governance Stack, which integrates six connected governance layers: permission, transparency, verification, human agency, institutional infrastructure, and accountability. The framework offers universities a practical socio-technical model for governing GenAI responsibly while protecting learning, privacy, academic integrity, and human oversight.

Chi-Square Feature Selection and SMOTE-Balanced Ensemble Learning for Explainable Cervical Cancer Risk Prediction

• A compact cervical cancer risk-prediction framework
using training-data-only Chi-square feature selection to
retain 27 informative features.
• A Soft Voting ensemble of Logistic Regression, XGBoost, and CatBoost combined with SMOTE-based classimbalance handling.
• A multi-level SHAP analysis providing global feature importance, feature directionality, and individual prediction
explanations.
• An interactive research prototype that unifies modelbased risk prediction, probability and threshold information, and feature-level explanations within a single
interface.

From Visibility to Action: A Multimodal Content Analysis of Digital Sustainability Communication in Bangladeshi Universities

The paper’s main contribution is identifying a “visibility-to-action gap” in Bangladeshi universities’ digital sustainability communication. It introduces the Message–Medium–Meaning–Action framework to show that effective communication must go beyond reporting institutional achievements by connecting credible evidence and sustainable technologies with clear participation opportunities, practical guidance, and accountability. Thus, it positions digital communication as a human-centered bridge between university sustainability systems and meaningful stakeholder action.

A Human-Centered and Sustainable Evaluation of Large Language Models: Empathy, Helpfulness, Cultural Sensitivity, and Safety

Large Language Models (LLMs) are increasingly used in contexts where interaction quality depends not only on technical performance but also on human-centered qualities. This study presents a user-centered comparative evaluation of four LLMs—ChatGPT, Gemini, Claude, and Grok—across empathy, helpfulness, cultural sensitivity, safety, and overall human-centeredness. A total of 112 participants evaluated 16 selected responses (four responses per human-centered dimension) drawn from a common set of 24 prompts; the full response pool contained 96 LLM-generated responses. Quantitative analysis was conducted using descriptive statistics and repeated-measures analysis of variance, supplemented by post-hoc comparisons and categorical preference frequencies. The results indicate that the models were generally perceived similarly across most dimensions. Before correction for multiple comparisons, empathy and overall human-centeredness showed the strongest unadjusted differences, with Gemini receiving higher ratings than ChatGPT; however, after Holm correction across the five outcome tests, no dimension reached statistical significance. Preference patterns nevertheless varied by dimension, with Gemini most frequently selected for empathy and overall response preference, while Claude was most frequently selected for helpfulness and trust. These findings suggest that user judgments may vary by dimension, supporting the value of multidimensional user-centered evaluation.

BriXNet: Efficient Depthwise-Separable CNN for Lightweight Brain Stroke Classification from CT Images

• We design a compact architecture containing only 27.91K
model parameters, built primarily from depthwiseseparable convolutions.
• We evaluate the proposed model on two public CT-image
datasets using accuracy, precision, sensitivity, specificity,
F1-score, Matthews correlation coefficient (MCC), and
AUC.
• We conduct within-dataset, cross-dataset, and five-fold
analyses, with explicit discussion of data leakage and
external-validity risks

Human-Centered AI for Sustainable EFL Speaking Development: A Reflective Learning Intervention in Bangladesh

Sustainable educational technology should extend meaningful learning opportunities without displacing teacher judgment, learner agency, or equitable participation. This mixed-methods study examined a semester-long, technology-enhanced reflective learning intervention for undergraduate English-speaking development at a private university in Bangladesh. The intervention combined repeated speaking tasks, video-based self-review, teacher feedback, peer feedback, guided self-assessment, and supplementary ChatGPT-mediated prompts. Thirty-five students completed matched pre- and post-intervention speaking assessments across fluency and coherence, grammatical accuracy, vocabulary resource, pronunciation and intelligibility, and interaction and effectiveness; four focus-group discussions (FGDs) explored perceived processes, benefits, and challenges. Overall scores rose from M = 18.63 (SD = 2.44) to M = 21.97 (SD = 2.35), t(34) = 18.37, p < .001, with a very large within-group effect (d = 3.11). All five dimensions improved significantly. Thematic analysis indicated that replayable performance evidence and human feedback supported error noticing, goal setting, repeated practice, confidence, and participation. ChatGPT was useful for vocabulary exploration, alternative expressions, planning, and low-stakes prompting, but memorizing AI-generated scripts could weaken spontaneity and increase dependence. Sustainability was not directly measured; rather, the findings support a bounded sustainability interpretation. A low-material, reusable reflective cycle can extend practice beyond class and reduce dependence on continual teacher-produced resources, but its long-term viability depends on device and connectivity access, AI literacy, privacy safeguards, teacher oversight, and proportionate use that recognizes generative AI’s environmental costs. The study contributes a human-centered model in which AI augments a socially grounded reflective ecology instead of replacing it.

Dual-Attention ResNet50-LSTM for Video-Based Human Action Recognition

Abstract—Human action recognition (HAR) from video requires effective modeling of both spatial appearance and temporal dynamics. Although convolutional neural networks (CNNs) can extract discriminative spatial features from individual frames, they are limited in modeling long-term temporal dependencies. Long Short-Term Memory (LSTM) networks address this limitation by learning sequential representations; however,
conventional CNN-LSTM frameworks may assign equal importance to irrelevant spatial regions and temporal observations. This paper proposes a dual-attention ResNet50-LSTM framework for video-based human action recognition. A pretrained ResNet50 network first extracts spatial feature maps from uniformly sampled video frames. A spatial attention mechanism then identifies informative regions within each frame, while an
attention gate regulates the attended features before sequential processing by an LSTMCell. Subsequently, a temporal attention mechanism assigns adaptive importance weights to hidden representations across the video sequence for final classification. The proposed framework is evaluated on the UCF50 and UCF101 benchmark datasets and achieves test accuracies of 94.10% and 92.42%, respectively. The results demonstrate the effectiveness of combining spatial and temporal attention with CNN-LSTM based
video representation learning.

The major contributions of this work are summarized as follows:
• A dual-attention ResNet50-LSTM framework is proposed for video-based human action recognition.
• A spatial attention mechanism is employed to emphasize discriminative regions within individual video frames.
• An attention-gating mechanism regulates the attended spatial features before sequential processing.
• A temporal attention mechanism is used to adaptively aggregate LSTM hidden representations.
• The framework is evaluated on UCF50 and UCF101, achieving test accuracies of 94.10% and 92.42%, respectively.

Rethinking Human-Centric AI Responses: Evaluating Support versus Constructive Challenge for Sustainable Human-AI Interaction

Human-AI interactions have become an integral part of daily life, influencing how people solve problems and make decisions. However, AI’s tendency to blindly agree with users can reinforce misconceptions, biases, and discourage critical thinking. While previous studies have focused mainly on AI performance and user satisfaction, less attention has been given to how constant AI support affects human thinking. This study examines whether human-centric AI should always support users’ perceptions or respectfully challenge them when necessary. A mixed-methods approach was used, combining prompt-response analysis with a survey measuring trust, perceived helpfulness, and critical thinking. The findings show that constant agreement can create an “echo chamber,” reinforcing personal biases, while respectful challenges encourage users to reconsider their beliefs and think more deeply. Three response types were identified: general agreement, constant support, and polite counterargument. The study concludes that human-centric AI should balance emotional support with constructive challenges to promote critical thinking and better decision-making.

Air Quality Forecasting for Dhaka: Data Integrity over Model Complexity

We show that data integrity and feature design dominate model complexity for Dhaka AQI forecasting. A two‑check audit reveals that a widely used 26‑year hourly dataset is synthetic before August 2022; training on it yields negative skill. On the verified record, LightGBM with perfect‑prognosis weather features significantly outperforms a Transformer fusion model at +24 h (Diebold–Mariano tested). An operational per‑horizon forecaster retains positive skill from +1 h to +72 h using only free public data.

Self-Supervised Contrastive Learning with Attention Refinement for Chest X-Ray Disease Classification

This work combines momentum-based contrastive self-supervised (MoCo v2) pretraining with a CBAM-augmented ResNet-50 classifier to investigate their joint and individual contributions to four-class chest disease classification. A controlled ablation isolates the effect of pretraining strategy by holding the architecture, loss function, and training schedule fixed between the self-supervised and ImageNet-initialized variants.

Technical Assessment of Electric Three-Wheeler Growth, Solar-based Charging and Grid Impacts

The rapid growth of battery-run electric three-wheelers (E3Ws) in Bangladesh creates emerging distribution-grid challenges due to uncontrolled fleet expansion and unmanaged charging practices. In this paper, we propose an integrated framework that couples E3W fleet forecasting, stochastic charging demand, solar-assisted charging, distribution-grid hosting capacity and policy assessment. A weighted logistic model using five published estimates of fleets estimates about 7.3 million E3Ws in 2030. Charging simulations using the Monte Carlo method suggest that uncontrolled charging can lead to an extra peak demand of approximately 5~GW in the evening. Six energy-management scenarios are compared on benchmark feeders based on measured solar-generation data, hosting-capacity, voltage performance, losses, cost, carbon-reduction, and renewable utilization, using linear-programming based scheduling. With the solar-assisted smart charging, the best overall balance is achieved with 77\% renewable charging, 69\% carbon reduction, and improved hosting capacity without battery storage. PV-plus-storage allows for more renewable integration, but is still not competitive at current values. The study also links the quantified technical effects to regulatory shortcomings and suggests an evidence-based roadmap with a focus on fleet registration, managed charging, charger standards, and integration of renewables. The findings show that using coordinated charging strategies can turn fast-growing E3W fleets into flexible distributed energy resources.

Fine-Tuned EfficientNetB0 for Pneumonia Detection from Chest X-Rays: Independent Evaluation and Grad-CAM-Based Explainability

Pneumonia remains one of the most common causes of hospitalization and a leading respiratory illness worldwide. Chest X-rays (CXR) are widely used to diagnose pneumonia, but their interpretation can vary from one doctor to another. This paper presents a fine-tuned EfficientNetB0 convolutional neural network for binary classification of pediatric chest radiographs into two categories: Normal and Pneumonia. Using the publicly available Kaggle Chest X-Ray Pneumonia dataset, a frozen-backbone transfer-learning baseline was first trained and then compared against another EfficientNetB0 model where the upper layers were fine-tuned using a low learning rate while the batch-normalization settings were kept unchanged. The frozen baseline collapsed to majority-class prediction, whereas fine-tuning produced strong, balanced validation performance (ROC-AUC 0.9923). On a fully independent, untouched test set of 624 images, the final model achieved 87.98% accuracy, 95.90% ROC-AUC, 97.44% sensitivity, 72.22% specificity and 91.02% F1-score at a decision threshold of 0.50. Threshold sensitivity analysis, per-class error analysis and Gradient-weighted Class Activation Mapping (Grad-CAM) visualizations for true-positive, true-negative, false-positive and false-negative cases are reported to assess whether the model attends to clinically plausible lung regions. Grad-CAM analysis provides additional visual insight into the image regions influencing model predictions and supports the interpretability of the classification results.

An Integrated Semantic Framework for Surveillance Video Information Retrieval

Modern surveillance systems continuously produce
large volumes of video, although much of the recorded footage
contains only a static background or very limited activity. Storing
such data can therefore become expensive, while finding a particular event in a long video sequence is often slow and difficult
to perform manually. To address this issue, we developed a
knowledge-based framework that transforms surveillance footage
into a structured and searchable representation. The framework
first applies a motion-based filtering step to identify frames
containing meaningful activity, which helps reduce the amount
of video processed by the object detector. The detected objects
are then tracked over time, and their movement patterns are
interpreted using an ontology to identify events such as walking,
running, falling, and intrusion. Information about the detected
events is represented as RDF triples and stored in a knowledge
graph, where it can be queried through SPARQL. We also
provide a natural-language interface that converts user questions
into SPARQL queries, making the event database accessible
without requiring users to have knowledge of the query language.
An evaluation using a subset of the VIRAT dataset shows that
the framework can detect and represent surveillance events
while reducing the amount of data that needs to be processed
and stored. The results also indicate that the natural-language
interface can retrieve the relevant events with good accuracy.

Rethinking Long-Term Evolution Throughput Prediction from Passive Drive Tests: Target Semantics and Dependence-Aware Validation

As seen in this research, passive phone bitrate logs cannot be taken as a sign of actual Long-Term Evolution (LTE) network capacity. The study compares conventional random validation with a combination of target-semantics analysis, dependence-aware validation, spatial holdout testing, and simple baseline comparisons, and demonstrates the superiority of the combined approach to avoid over-estimating the performance of machine-learning models in drive-test data. The main contribution is a more reliable framework for evaluation of cellular throughput prediction models based on passive mobile measurements.

ValeLink: Design, Development, and Evaluation of a Digital Property Management Platform for Fiji’s Residential Rental Sector

The research illustrates how digital technologies can modernize rental management and improve service delivery in small island developing countries that are still evolving. The research is also of interest in that it adds to the growing literature on PropTech adoption and provides a case of its implementation and evaluation in Fiji. The results also demonstrate that in order to achieve digital transformation in the residential rentals industry it is important not only to have the technological capabilities, but also to meet local regulatory, cultural and operational requirements.

CitizenLink: Scalable Graph-Based Modeling and Analysis of Multi-Generational Citizen Relationships

This paper contributes a graph database approach to Bangladesh’s national citizen record keeping. It uses Neo4j to store citizens as nodes and family ties as direct links. This lets the system answer deep family queries that relational databases handle poorly. The work delivers three main things. First, it builds a realistic synthetic population of over 5 million citizens across four generations. The generator follows 47 rules drawn from real Bangladesh data on marriage age, religion, fertility, and polygyny. Since the data is synthetic, others can reuse it without privacy risk. Second, it proves the speed claim at scale. Indexed lookups run in about 13 milliseconds. Relationship queries stay under 100 milliseconds on a graph of 15.3 million elements. Query cost depends on family size, not population size. Third, it introduces a relationship classifier that reads the order of steps along a path. Simple hop counting cannot separate a co-wife from a step-mother, because both sit two hops away. This classifier gets both right. It also separates half-siblings from full siblings using a parents map. The system further adds a rule based birth audit that flags biologically impossible birth gaps and duplicate registrations. The rules are open and checkable, so officials can justify any decision. Earlier AI methods were accurate but gave no auditable reasoning. This paper closes that gap by joining graph speed with legal transparency in one working prototype.

Token Smuggling and Log Injection in Open-Source AI Coding Assistants: A Case Study

Our contributions are as follows.
– We build a controlled harness that plants benign canary payloads in build output read by an open-source coding assistant, grades the assistant’s response on a four-state scale, and records every trial for reuse.
– We show that payload encoding drives success independently of wording. Carriage-return injection holds the attack at 80.0% while concealing it from a human reader; zero-width concealment drops it to 13.3% by breaking tokenization, with a non-deterministic residual that rules it out as a defense.
– We show that the assistant’s built-in filter keys on register: an overt directive is refused in every trial (0%), while the same instruction written as a build notice runs in 66.7%, and disabling confirmation lifts that only to 86.7%.
– We test two defenses and find that stripping control characters does nothing to the plain-text attack, while checking the log against the real build output stops it completely.

Robust Bangla Sarcasm Detection with HFAT-v4: Multi-Seed Evaluation, Ensemble Learning and Supporting English Benchmark Analysis

Automatic sarcasm detection remains challenging in low-resource
languages because pragmatic intent is often implicit, context dependent and
weakly represented by surface-level cues. Existing Bangla studies have
largely adopted binary formulations and have reported limited evidence on
stochastic stability, class-specific reliability, confidence calibration and
systematic errors. This study introduces HFAT-v4, a BanglaBERT-based
Hybrid Fourier-Attention Transformer for three-class classification of
Neutral, Non-Sarcastic and Sarcastic text. HFAT-v4 combines scalar-mixed
contextual representations with self-attention, adaptive Fourier token mixing,
a spectral residual pathway, multi-representation pooling, hierarchical
supervision and R-Drop regularization. Following normalization, conflicting
duplicate removal and same-label deduplication, 11,911 Bangla instances
were stratified into 8,337 training, 1,787 validation and 1,787 test samples.
Performance was evaluated across five random seeds and three ensemble
strategies using classification, ranking, calibration and error-analysis
measures. The five-seed mean accuracy and macro-F1 were 0.6384 ± 0.0047
and 0.6386 ± 0.0038, respectively. Majority voting achieved the strongest
hard-label performance, with 0.6575 accuracy and 0.6584 macro-F1, whereas
probability averaging achieved the best macro-ROC-AUC (0.8254), macro
average precision (0.7068) and expected calibration error (0.0795). Sarcastic
text was recognized most reliably, while non-sarcastic text remained the
principal bottleneck; 376 test instances were misclassified by all five models,
indicating systematic ambiguity beyond seed-specific variation. A
complementary English Reddit analysis also showed gains from richer n
grams and subreddit context. These findings demonstrate the value of multi
seed and calibration-aware evaluation while identifying contextual ambiguity
as the primary target for further improvement.

Beyond Ground Truth: A Heuristic Unsupervised Quality Assurance Engine for Diagnosing Calibrational Blindness in Dental Segmentation

This research proposes a novel heuristic unsupervised quality assurance engine designed to detect and diagnose calibrational blindness in dental segmentation models without relying strictly on ground truth. By addressing the limitations of traditional verification methods, the study introduces an automated evaluation framework developed specifically for dental radiography. It also shows how the framework improves the reliability, transparency, and practical applicability of AI-driven dental diagnostic support tools for real-world clinical use.

Cross-Generational Transportability of Cognitive Risk Models: A SHAP-Based Analysis from Aging Adults to University Students

This study evaluates the portability of cognitive risk models across generations by training a LightGBM classifier on 65,914 aging adults from the LASI dataset (age 45+) and testing the trained algorithm on 297 university students (age 18-30). While the initial direct transfer only achieved an AUC of 0.586 due to distribution shifts, applying Platt scaling calibration and threshold optimization improved performance, achieving an AUC of 0.672 and an F1-score of 0.529, which is better than the baseline model. Moreover, the analysis revealed that the calibration of probability bridges the performance gap between different generations, while the domain shift from objective cognitive testing in the elderly to subjective cognitive complaints in students accounts for most of the performance drop.

Trustworthy Screening of CES-D-Defined Elevated Depressive Symptomatology Through Evidence-Guided Ensemble Selection and Independent Structural Validation

Depression-severity classifiers built on the Center for Epidemiologic Studies Depression Scale (CES-D) are usually evaluated on accuracy alone, without checking whether the 20item instrument behaves coherently on the population studied. We address both questions on 896 Bangladeshi university students. First, a bank of 12 classifiers is benchmarked on mapping the 20 CES-D items to four severity categories; leave-one-out ablation and forward selection identify a two-member ensemble of Logistic Regression and TabPFN, a pre-trained tabular foundation model applied zero-shot. On a locked test set, this ensemble reaches ROC-AUC 1.000, balanced accuracy 0.986, and F1 0.986 for binary elevated-symptomatology classification, and balanced accuracy 0.983 for the four-class task. Because the severity label is a deterministic function of the same 20 items, this is reported as classification-rule recovery rather than independent prediction. Second, the CES-D’s internal structure is validated using itemnetwork predictability– each item modeled from the other 19, with no circular dependency on the target– cross-checked against exploratory factor analysis (KMO = 0.929, Cronbach’s α = 0.852). Both methods converge on the same core symptom cluster (felt sad, felt depressed, loneliness), consistent with the standard four-factor CES-D structure. Calibration, subgroup fairness, and SHAP and permutation-importance explainability are reported for the ensemble as a unit.

Comparative Analysis of Class-Weighted Random Forest, XGBoost, and SVM for Multi-Class Photovoltaic Fault Classification

This research develops a machine learning-based approach for accurate fault classification in photovoltaic (PV) systems using Random Forest, XGBoost, and Support Vector Machine (SVM). The study compares the performance of these models to identify an effective and reliable approach for detecting and classifying PV system faults.

A Techno-Economic Framework for Divisional Renewable Energy Cost Analysis across Bangladesh’s Grid Zones Using Machine Learning

Bangladesh aims to source 20% of its installed generating capacity from renewable energy by 2030, yet the delivered cost of renewable electricity across its eight administrative divisions has never been established on a common basis. This study
develops a zone-wise levelized cost of energy (LCOE) framework that integrates division-specific distribution loss rates and the regulated retail tariff with solar irradiance and wind speed, and maps the resulting cost surface using regression-based sensitivity analysis. A discounted cash flow model with a delivered-energy correction is parameterised from verified sources and evaluated for utility-scale solar photovoltaics (PV) and onshore wind in each division. Three machine-learning regression models – linear, polynomial and random forest – are trained on Latin-hypercube samples and compared against one-at-a-time sensitivity analysis.
Solar PV LCOE is consistent across the eight divisions (0.092–0.099 USD/kWh) and reaches near-parity with the tariff in four of them, while onshore wind is competitive only in the three coastal divisions and is costlier than PV everywhere. Divisional irradiance varies by only 7.7%, whereas a 2% point reduction in the cost of capital lowers LCOE by 13.2%. The results show that financing conditions, not resource geography, constrain renewable cost competitiveness in Bangladesh.

Quantifying Social Bias in LLMs for Bangla: A Likelihood-Based Comparative Analysis

This study provides a comparative evaluation of stereotypical social bias in three instruction-tuned LLMs on the Bangla BanStereoSet benchmark across nine bias categories. A key contribution is the systematic comparison of whole-sentence and candidate-token-conditioned likelihood scoring, demonstrating that the scoring methodology can substantially alter measured bias. The findings further show that Bangla-language specialization alone does not guarantee reduced stereotypical bias, highlighting the need for language- and script-aware fairness evaluation for Bangla LLMs.

A Cross-Dataset Robustness Benchmark for Driver Drowsiness Detection: Quantifying and Mitigating the Generalization Gap

The paper’s central contribution is showing that the near-perfect accuracy these drowsiness-detection architectures report is largely an artifact of testing on the same dataset they were trained on: both models lose a real, measurable fraction of their accuracy the moment they’re evaluated zero-shot on an independent dataset (custom CNN 99.7%→82.35%, transfer learning 99.1%→87.65%), and you introduce the Generalization Ratio as a simple, reusable way to express that drop against each architecture’s own published ceiling.

The more significant methodological finding, though, is what that gap turned out to actually be made of. A prior single-run reproduction of this same protocol found a much larger gap and attributed it to an architectural limitation — but by repeating every condition five times, you discovered that two of five seeds had simply collapsed from a bad weight initialization, not a genuine ceiling on generalization. Fixing that (He-init) and re-averaging shows the real gap is smaller than previously reported, meaning part of what looked like “domain shift” in earlier work was actually undiagnosed training instability. That’s a genuinely useful correction for the field, not just your own paper.

The other concrete contribution reviewers tend to value: the condition-stratified analysis pins the remaining gap on sensor identity specifically (not eyewear or lighting), and the error asymmetry (models fail safe, over-predicting “closed” rather than missing drowsiness) is a practically meaningful result for anyone deploying this class of system.

Privacy-Preserving Federated Learning for ICU Mortality Prediction in Hypotensive Patients: Differentially Private Federated Learning with a Personalization Ablation

The main contributions of this work are the development of a hypotension-focused ICU cohort from MIMIC-IV v3.1, the integration of GRU-D with DP-SGD and Per-FedAvg-style personalization, and a comprehensive comparison with centralized, federated, privacy-preserving, local-only, and SOFA-based approaches. The ablation study shows that personalization provides no statistically significant benefit under non-IID and privacy constraints. Additionally, severity and feature-importance analyses identify Glasgow Coma Scale (GCS) as the strongest predictor of mortality in the cohort.

Climate Variability and Rice Production: Extreme Temperature Index Modeling: Seasonal Yield Response Assessment

It creates a novel dataset that explicitly separates weather conditions (heatwaves, rainfall, humidity, sunshine) in the flowering and grain filling phases simultaneously for all three rice seasons (Aus, Aman, Boro) in Bangladesh, filling a critical gap in weather stress studies.

Strong Methodological Framework: It presents a custom-made Heat Stress Index (HSI) as well as a consensus feature-selection scheme (Pearson-MI-RFE) and a Non-Negative Least Squares (NNLS) stacking ensemble. The ensemble (chronological splitting (leakage free) and statistical tests (Friedman, Wilcoxon, DM) provides a highly accurate prediction,

In terms of Actionable Climate Risk Quantification, it presents for the first time a SHAP-driven seasonal vulnerability ranking (Aus > Aman > Boro) and quantifies specific yields loss projections under IPCC warming scenarios (+3°C: 43.1% loss for Aus). It provides a concrete, data-driven basis for targeted climate-adaptation strategies in rice production.

A Modular Distributed IoT Architecture for Closed-Loop Hydroponic Monitoring and Control with Image-Based Plant Observation

The main unique contribution of the research reported is a modular, distributed IoT based hydroponic architecture that combines different functions together, rather than treating sensing, control and planting observation as distinct subsystems. The system consists of a Raspberry Pi supervisory controller with distributed ESP32 field nodes, providing multi-parameter sensing and closed-loop control to the parameters such as nutrient dosing, pH, water circulation, humidity, ventilation and lighting in a callable manner. It also contains a first time non-contact plant segmentation, projected plant area, and colour-based features with camera usage as well as MQTT/Modbus communication and persistent data and remote visualisation. End-to-end integration including sensing, communication and monitoring, logging and actuator control verified in prototype.

Design and Development of a Low-Cost IoT-Based Smart Kitchen Safety and Automation System

The key contribution of this paper is KitchenGuard, a low-cost smart kitchen system based on ESP32 that goes beyond traditional hazard monitoring by enabling detection, autonomous decision-making and immediate active physical mitigation. In contrast to many existing systems that merely communicate gas detection and alert the users, this paper presented KitchenGuard which features integration of gas-leakage, water-leakage, and motion-sensing with smart automatic gas/water valve isolation, exhaust turn-on/turn-off, alarm notification porting system as well as occupancy-based lighting control in one interoperable architecture.

ResQWatch: A Bluetooth-Enabled App Based Wearable GPS-Aware Extension for Rapid Emergency Trigger and Automated Police Dispatch.

This work’s contributions, briefly:
1) A debounce-checked double-press gesture with an explicit timing-window model, so accidental single knocks
and switch bounce don’t get treated as real triggers.
2) An auto-reconnecting BLE 5.0 relay that hands a distress
record from a phone-independent wearable to a bonded
companion app, without the onboard GPS/GSM hardware that dominates cost and power draw in most prior
designs.
3) A cloud dispatch scheme that runs geofenced community notification and automated nearest-station police
contact at the same time, using Haversine-distance POI
lookup.
4) A full application stack, React Native app, .NET services, PostgreSQL, plus everyday features (a social
safety feed, a government notice board) meant to keep
the app open on ordinary days, not just emergencies.
5) A theoretical, stage-by-stage latency budget for the
whole press-to-dispatch chain, built from BLE spec timing and datasheet figures presented honestly as design
targets, not measured numbers.

AMIGO: An IoT-Enabled Autonomous Mobile Robot for RFID-Verified Hospital Medicine Delivery with Cloud-Integrated Fleet Monitoring

The contributions of this paper are:
1) A low-cost delivery robot combining Arduino control,
ESP8266 connectivity, line-following navigation, obstacle detection, and RFID-based destination identification.
2) An RFID-to-cloud-to-web pipeline that links a physical delivery event to patient, bed, ward, and medicine
records, validated through repeated-trial testing of RFID
recognition, line-following navigation, and end-to-end
web synchronization.
3) A web dashboard, backed by Supabase, for hospital staff
to monitor deliveries, RFID scans, and robot status in
real time.

Ordinal Annotation of Pre-Harvest Rose Maturity in Field Collected Images: Inter-Annotator Reliability and Feature-Space Analysis

The main contributions are: 1) a morphology-based annotation protocol for
four ordered pre-harvest maturity stages; 2) a protocol-annotated field
dataset of 681 individual rose images; 3) an inter-annotator reliability
assessment on a blinded, stratified 50-image subset; and 4) a multi-metric
feature-space analysis characterizing global structure and local ordinal
consistency without supervised training on the proposed labels.

An Explainable Metaheuristic Optimized Feature Engineering Framework for Enhanced Heart Disease Prediction

The study contributes a clinically informed feature engineering strategy, a hybrid MI–mRMR–PCA feature optimization approach, GA-based model tuning, and comprehensive model evaluation with SHAP-based explainability.

A Hybrid EfficientNetB0–XGBoost Framework for Betel Leaf Health Classification Using Field Images

• Self-Collected Dataset: A real-world Piper betle leaf
disease dataset personally collected under natural field
conditions in Rajshahi, Bangladesh.
• Hybrid Framework: An EfficientNetB0–XGBoost
framework for betel leaf disease classification.
• Feature Selection: XGBoost-based selection of informa-
tive deep features for efficient classification.
• Optimal Representation: Identification of an effective
feature representation for improved classification perfor-
mance.
• Comparative Analysis of ML Models: Comparative
evaluation of multiple machine learning models to iden-
tify the most effective model for disease classification.

Fabrication of Jute Latex Laminated Composite for Packaging Material

This study examined the multidirectional mechanical, wetting, and thermal performance of jute latex laminated composite (JLLC) developed for cereal material. Woven jute fabric was laminated with natural rubber (NR) latex at four different loadings (1 to 4 g/100 mL toluene), while unlaminated jute fabric was used as the control. The bursting strength, bursting time, wetting time, and thermal resistance of the developed composites were evaluated. All laminated samples exhibited higher bursting strength than the control fabric, with JLLC1 showing the highest value of 11.10 kg/cm², representing a 35.9% improvement over the control. Bursting time increased from 22.65 s for the control to 25.29 s for JLLC4. The latex layer substantially improved water resistance; while the control fabric wetted from the bottom side within 2.15 s, all JLLC samples remained unwetted for the 120 s observation period. Thermal resistance also increased from 0.02374 to 0.05833 m² K/W, with JLLC4 showing the highest value. Overall, NR latex lamination effectively enhanced the functional performance of jute fabric, demonstrating its potential for moisture-resistant and thermally resistant packaging materials.

Explainable Machine Learning for Regional Load-Shedding Severity Classification Using Supply-Side Constraints in Bangladesh

An explainable three-class severity classification framework is proposed for nine divisions based on generation, supply-side constraints, weather data, calendar data, and regional characteristics. Various classifiers and a probabilistic soft Voting Ensemble approach are considered for severity classification in case of regional class imbalance. SHAP explanation provides insight into significant features and distinguishes the role of national generation versus regional supply-side constraints. A regional vulnerability analysis combines model explanations with observed load-shedding events to provide division-specific operational and policy insights.

Explainable AI Based Dual-Domain Deep Network for Brain Tumor Classification

The principal contributions of this work are summarized as follows:
• A dual-domain fusion architecture that couples a frozen EfficientNetB0 spatial encoder with a purpose-built wavelet-domain CNN, addressing the exclusively spatial domain limitation shared by recent approaches.

• An explicit four-subband wavelet feature-engineering pipeline in which the LL,LH,HL,HH decomposition is presented as a four-channel tensor, enabling the network to learn inter-subband correlations — such as the co-occurrence of horizontal and diagonal detail energy that characterizes a spiculated infiltrative margin—that a spatial-only CAM-interpretable model cannot access directly.

• A dual-modality explainability protocol integrating Grad-CAM and a dual-input-faithful adaptation of LIME, extending the CAM-only interpretability practice, independent attribution method and explicit input-space faithfulness for the frequency branch

D3QN-Driven Cooperative Partial Task Offloading for Latency-Energy Optimization in 6G LEO Satellite-Terrestrial Edge Networks

This work proposes CoPTO (Cooperative Partial Task Offloading), a three-segment LEO satellite-terrestrial edge computing framework that, unlike prior approaches which only choose between a single local device/edge server or an all-or-nothing satellite offload, enables a divisible computation task to be dynamically split for parallel execution between a serving terrestrial edge server (TES) and a neighboring TES — while retaining full satellite edge server (SES) offloading as an alternative when terrestrial conditions are unfavorable. This exploits previously unused spare capacity at neighboring terrestrial servers, which no existing DRL-based offloading scheme specifically targets.

A Dueling Double Deep Q-Network (D3QN)-based Intelligent Supervisor learns online to choose between full SES offloading and a continuous TES-TES split ratio, jointly minimizing latency and energy under per-task deadline and energy-budget constraints — a decision problem that is non-convex and time-varying, making it unsuitable for static optimization.

Simulation results show the approach reduces average normalized latency by up to 22.45% and energy consumption by up to 20.62% compared to a non-cooperative D3QN baseline, with the gains growing as task load increases (due to better mitigation of queue build-up at heavily loaded single servers).

SegSwin-Net: A Segmentation-Guided Vision Transformer for Improved Leukemia Detection and Classification

This study introduces a novel hybrid framework that enhances automated leukemia detection by using cell segmentation to eliminate background noise prior to Swin Transformer classification. By isolating these key diagnostic features, the proposed model achieves an exceptional 99% accuracy, significantly outperforming baseline CNN architectures.

A Trusted Threat Intelligence Acquisition Framework for Water Treatment Cyber Physical System

Recently water treatment facilities (WTF) equipped their cyber physical system (CPS) with latest mechanism which make the system capable for superior assistant for the water treatment operations. On the other hand, this mechanism has drawbacks which make the WTF CPS porn to cyber physical assaults. So, proactive cyber defense method is required for protecting WTF CPS from cyber-attacks. For this purpose, cyber threat intelligence (CTI) is a valuable resource for designing effective defense systems for WTF CPS. Latest threat information is perpetually accommodated by CTI which play pivotal role for architect and deploy security initiatives and alleviation policies. CTI collection from different sources and validation is important for ensuring the quality of CTI. CTI from external sources may contain fake CTI. In this paper we proposed a CTI gathering framework for WTF CPS combining both active and passive sources. We also designed a validation and scoring algorithm for those gathered CTI.

SCSRF: A Stable Clinical Survival Risk Framework for PBC Survival-Risk Prediction

Predicting likelihood of death in primary biliary cholangitis (PBC) is unreliable must be handled with great care for small samples, censoring, model selection, etc. Predictor stability this study propose the following Stable Clinical Survival Risk Framework. The model was built using (SCSRF), which combines the local feature engineering technique of folding with the cross-validation technique. Evaluation that takes censorship into account, and stability analysis. Among three learners, al
though the random survival forest’s mean C-index value (0.855) was the highest, the differences between the pairs were not significant post Holm correction. Its pooled out-of-fold C-index was 0.832 (95% CI: 0.800–0.862), with an IPCW C-index AUC
of 0.876, time-dependent AUC of 0.788 and integrated Brier score of 0.117. Even with just 15 features, the model’s ability to predict survival risk was similar to using all the features (C index = 0.830). The curves of the out-of-fold risk groups were
similar when plotted with the Kaplan–Meier method. There was consistent separation of risk groups across three repeats (median log-rank p = 2.42 × 10−28). Internal transportability analysis yielded a C-index of 0.786. These findings indicate that SCSRF is a a framework for survival-risk stratification based on reliability; But it must be independently externally validated to make it generalisability and clinical usefulness.

Explainable and Generalizable Deepfake Detection for Vishing Attack Recognition

Artificial Intelligence (AI) has come a long way in
mimicking human voice to the extent that it cannot be identified
without extended effort. As a result, AI-generated vishing attacks
have increased threat, as humans mostly get deceived by the
highly human-alike AI-generated voice. Proposed study aims to
build a generalizable and explainable framework that can identify
both deepfake and scam intent. For deepfake detection, a voting
ensemble containing Hubert-base and WavLM-large is zero-shot
tested on the out-of-domain for generalization. Additionally, for
scam-intent classification, DistilBERT classifies scam intents for
transcripts generated via Whisper speech-to-text. Both branches
produce a combined result, which can be from three classes: AI-
Legit, AI-Scam, and Human-Legit, utilizing a rule-based joint
probability threshold of 0.5. Shortcut learning is evaluated using
Generalization, Robustness, and Stability Index (GRSI), a six-test
perturbation battery, and silence diagnostics. Dual-level SHAP
is used to provide acoustic and word-level explanation. During
held-out internal testing, DistilBERT showed 99% accuracy with
0.98 recall, while the ensemble for deepfake detection showed
90% accuracy after soft voting. Hence, the proposed framework
showcases generalization, and explainability in identifying AI-
enabled attacks while decreasing the risk of shortcut learning

A Sequential Hybrid Intrusion Detection Model for Detecting DoS Attacks in UAV Networks

The significant research contribution is the design of a sequential hybrid IDS (MLP with Random Forest) for detecting DoS attacks in UAV networks. The model combines the MLP’s ability to learn complex nonlinear traffic patterns with Random Forest’s ensemble-based decision refinement. Using 5-fold out-of-fold stacking and SMOTE-based class balancing, the proposed approach achieved 0.92 accuracy, 0.95 recall, 0.84 F1-score, and approximately 0.935 ROC-AUC, outperforming the evaluated standalone baseline models.

A Human-Centered Evaluation of AI-Generated Responses for Cultural Sensitivity in Large Language Models

The use of Large Language Models (LLMs) in culturally diverse settings has grown, but their capacity to produce culturally relevant output is not fully understood at the human level. There is a need to develop an AI framework that can evaluate responses for cultural sensitivity to empower non-Human Resource professionals to perform this task. Two dimensions of cultural sensitivity in AI-generated responses are proposed and evaluated in this study, cultural respect and contextual appropriateness, in an AI framework. 121 participants were recruited for the quantitative evaluation, which involved rating three different response blocks generated by AI on a five-point Likert scale. The two indicators were combined to create an overall measure of perceived cultural sensitivity, the proposed Cultural Sensitivity Score (CSS). Overall, the perception of the AI responses was positive with a mean CSS of 3.652 (SD = 0.869), well above the neutral value of 5 (t(120) = 8.247, p < .001). There were no significant differences between the three response blocks, suggesting equivalent perceived cultural sensitivity. Additionally, cultural respect and contextual appropriateness had a significant and positive correlation (r = 0.883, p < .001). The results demonstrate the value of human-centered evaluation to better understand culturally appropriate AI behavior and the need for cross-cultural studies and more comprehensive evaluation frameworks.

When Humor Hides Harm: A Fusion-Based Approach to Decoding Sentiment in Bengali Memes

Novel Dataset: A manually curated and annotated dataset of 3,000 Bengali social media memes, labeled for both sentiment (Offensive/Non-Offensive) and thematic category (Entertainment, Sports, Religion, Politics, Others) — a valuable resource for a low-resource language where such multimodal datasets are scarce.
Multimodal Fusion Framework: An OR-operation-based decision-level fusion approach that combines textual predictions (SVM, LR, RF, BERT) with visual predictions (ResNet50, VGG19) to jointly determine offensive content, demonstrating that multimodal integration outperforms unimodal approaches.
Comparative Model Evaluation: Systematic benchmarking of eight text–image fusion combinations, establishing that BERT+VGG19 achieves the highest performance (77% accuracy), and showing VGG19 consistently outperforms ResNet50 as the visual backbone across all pairings.
Thematic Categorization: Introduction of a five-category content classification scheme (beyond simple polarity labels), enabling more nuanced, culturally grounded analysis of Bengali meme content — an angle largely unexplored in prior Bengali sentiment analysis work.
Contribution to Low-Resource NLP: Advances content moderation tooling for Bengali, a linguistically and culturally underrepresented language in multimodal sentiment analysis research, addressing a real gap relative to well-resourced languages like English.

Does Cross-Encoder Reranking Close the Multi-Hop Performance Gap Between Knowledge Graph Retrieval and Vector RAG in Renewable Energy Domain QA?

Graph RAG is widely assumed to outperform vector retrieval on multi-hop questions, yet most graph pipelines retrieve an unordered candidate set and truncate it, discarding the bridging chunk that a multi-hop question depends on. This work isolates that failure mode. Three pipelines are built over a single knowledge graph such that consecutive pipelines differ in exactly one variable: the reranked variant differs from the unranked graph baseline only in scoring the full candidate pool before truncation rather than after. This separates ordering-before-truncation from cross-encoding itself, an ablation absent from published work, which applies reranking only to candidate lists that already arrive ranked. Single-hop and multi-hop results are reported separately with bootstrap confidence intervals and paired Wilcoxon signed-rank tests. Reranking closes the multi-hop gap to statistical indistinguishability but leaves a significant single-hop gap, and the gain costs 1.8 s of local compute with no additional model calls

Demand-Coupled Assessment of Rooftop Solar PV Integration for Dhaka MRT Line-6

Rapid urbanization and limited land availability increase the importance of utilizing existing transport infrastructure for renewable-energy deployment. The previous study has determined the technical feasibility of integrating PV (Photovoltaic) along Dhaka MRT Line-6, however, the estimated PV generation and real electricity demand of MRT needs to be better interpreted. This research proposes a PVsyst based station-group framework for the Uttara North-Motijheel corridor comprising 16 stations and correlates rooftop PV generation with the traction and auxiliary electric powers needed for battery-free grid-connected operation. Three representative station archetypes are modeled and scaled throughout the corridor, leading to an installed capacity of approximately 15.79~MWp and an average PV generation of 62.71~MWh/day, or approximately 22.89~GWh/year. The base case traction-energy requirement is estimated to be 54.87~MWh/day, yielding an overall PV-to-traction energy ratio of 114.3\%. This is not to say that this is a real-time solar penetration figure, as there would be a need for further grid balancing due to mismatch in PV generation and MRT demand, as well as auxiliary loads. If one assumes a total electricity demand of the houses in the MRT, the PV generation estimated in this study amounts to about 60-65\% of the annual electricity consumption of the rooftops. The study thus shows that the PV potential at the corridor scale should be assessed not only in terms of annual PV generation but also in relation to the full MRT load boundary and grid connected operating conditions.

L2L-Pix: A Training-Free Pixel Replacement Method to Mitigate Unlabeled Noise in Patch-Based Hyperspectral Image Classification

The main contributions of this paper are summarized as follows:

1. We identify the problem of unlabeled-pixel noise within cropped HSI test patches, an issue largely overlooked in prior patch-based HSIC literature.

2. We propose L2L-Pix, a lightweight, training-free algorithm that replaces unlabeled pixels within a patch with spectrally consistent labeled pixels and integrates directly into any patch-based classification pipeline, requiring no additional learnable parameters and negligible computational overhead.

3. We conduct extensive experiments showing that incorporating L2L-Pix consistently improves classification accuracy across multiple datasets and backbone networks.

Archives: Paper Submission