IUT Institutional Repository

IUT Institutional Repository is a digital service that collects, preserves, and distributes digital material. Repositories are important tools for preserving an organization's legacy; they facilitate digital preservation and scholarly communication.

Communities in IUT Institutional Repository

Select a community to browse its collections.

Now showing 1 - 5 of 12

Recent Submissions

Av Machine Learning-Based Lightweight Consensus Framework for Blockchain-Enabled IoT Systems
(Department of Computer Science and Engineering(CSE), Islamic University of Technology(IUT), Board Bazar, Gazipur-1704, Bangladesh, 2025-10-25) Dihan, Mubtasim Kamal; Abdullah; Amina
The integration of blockchain with the Internet of Things (IoT) offers a promising ap proach to address the limitations of centralized IoT architectures. However, existing blockchain consensus mechanisms are often unsuitable for resource-constrained IoT devices and dynamic network conditions due to high computational demands or re liance on monetary-based participant selection.. In this work, we propose Dynamic& Periodic Proof of Evolutionary Model (DP-POEM), a lightweight, machine learning based consensus mechanism tailored for blockchain-based IoT systems. DP-POEM selects a group of block producers for defined periods using supervised learning, re ducing redundant computation while ensuring security through randomized block productionwithinthegroup. Itincorporatesbothstaticanddynamicallychangingde vice features, such as battery level and CPU usage, to select capable nodes efficiently, and implements fair participation mechanisms to balance network involvement over time. Theoretical analysis and experimental evaluation demonstrate that DP-POEM achieves high scalability, low latency, enhanced security, and improved applicability compared to traditional and state-of-the-art consensus protocols in dynamic IoT en vironments.
A Hybrid Machine Learning and Deep Learning Framework for ECG-Based Cardiac Abnormality Detection
(Department of Computer Science and Engineering(CSE), Islamic University of Technology(IUT), Board Bazar, Gazipur-1704, Bangladesh, 2025-10-25) Khan, Dayan Ahmed; Ferdous, Zannatul; Jarif, Kazi Suny Al
Cardiovascular diseases remain a leading cause of global mortality, underscoring the need for reliable and efficient diagnostic tools. Thisthesis investigates machine learn ing and deep learning approaches for detecting cardiac abnormalities using electro cardiogram (ECG) data. Initial experiments explored classical heart disease datasets; however, due to their limited scale and clinical relevance, the research pivoted to the PTB-XL+dataset,whichprovidestabularECGfeaturesextractedfrom12-leadrecord ings. Acomprehensive set of baseline models was implemented, includingtraditional clas sifiers (Logistic Regression, Support Vector Machine, Random Forest, Naïve Bayes, XGBoost, LightGBM, CatBoost) and deep learning architectures (Shallow MLP, MLP, TabNet, ResMLP,DANet,AutoInt). ModelswereevaluatedonthepredefinedPTB-XL folds using accuracy, precision, recall, F1-score, ROC-AUC, and confusion matrices. The best-performing baseline models were gradient boosting ensembles (XGBoost, LightGBM,CatBoost)andselectedneuralnetworks(MLP,DANet,ResMLP).Building upon these results, a stacked meta-learning framework was employed, with Logistic Regression, LightGBM, and Shallow MLP as meta learners. Two strategies were ex plored: (i) usingmetafeaturesonlyand(ii)combiningmetafeatureswithoriginalfea tures. The hybrid approach achieved the highest performance, with LightGBM_Full attaining an accuracy of87.7%,F1-scoreof0.876, andROC-AUCof0.947,outperform ing individual baselines. These findings demonstrate the effectiveness of stacking ensembles for ECG-based classification, particularly when integrating meta features with original features. The research contributes a systematic benchmarking of models on PTB-XL+, highlights the superiority of boosting-based ensembles for tabular ECG features, and establishes a reproducible framework that can guide future clinical decision support systems.
Enhancing Transferability using Federated Learning for ML-based Network Intrusion Detection
(Department of Computer Science and Engineering(CSE), Islamic University of Technology(IUT), Board Bazar, Gazipur-1704, Bangladesh, 2025-10-25) Salvi, Rezwan Islam; Hafiz, MD RafinJawad; Kalam, Abeer Mahmood
Network Intrusion Detection Systems (NIDS) are critical for securing network-connected devices, yet their ability to detect novel attacks remains limited due to poor transferability across attack classes. This thesis explores federated learning (FL) to enhance transferability in NIDS, addressing privacy and data heterogeneity challenges in cross-device settings. Using the CIC-IDS 2017 dataset, we implemented FedAdam and Fedprox, two different Federated Learning Techniques, toimproveconvergenceoverthebaselineFedAvgalgorithm in a federated setup. Experiments revealed that FedAdam and FedProx yield improvements in validation accuracy. We also used Synthetic Minority Oversampling Technique (SMOTE) other thanBootstrapping for classbalancing, simulaterealisticvariationsandimprovemodel robustness. These methods aim to better tackle non-IID data and class imbalance, critical for transferability. This study underscores the potential of FL in NIDS, offering a foundation for developing models that can generalize across diverse attack classes while preserving privacy, with future work focusing on optimizing adaptive methods and integrating more diverse datasets.
Spectra Net: Hybrid Time and Frequency-Domain Modeling for Sustainable Cloud CPU Prediction
(Department of Computer Science and Engineering(CSE), Islamic University of Technology(IUT), Board Bazar, Gazipur-1704, Bangladesh, 2025-10-25) Siddiqui, Nahin F.; Hoque, Zarif Safwan; Haque, Md. Ehsanul
The proliferation of cloud computing has created an urgent need for sustainable resource management strategies to mitigate the growing energy consumption and carbon footprint of data centers. Accurate CPU workload forecasting is a cornerstone of this effort, en abling proactive resource allocation that minimizes energy waste from over-provisioning, and prevents performance degradation from under-provisioning. This study introduces SpectraNet, a lightweight hybrid model that advances sustainable cloud computing by delivering highly accurate CPU usage predictions with minimal computational overhead. By integrating time domain and frequency-domain analysis, SpectraNet effectively cap tures both transient and periodic workload patterns. Experimental results on the Azure VM CPU Readings and Alibaba Cloud Workload datasets demonstrate that this dual-domain approach significantly improves forecasting accuracy. SpectraNet demonstrates an com petitive balance of performance and efficiency, achieving a Mean Absolute Error (MAE) of 0.0549 on long-range forecasts. Critically, the model achieves this accuracy while being up to 19 times smaller (with only 89,805 parameters) and 5 times faster at inference than larger, more complex architectures. The model’s efficiency and small footprint make it a practical and scalable solution for real world, resource-constrained cloud environments. By enabling more precise and energy-efficient resource management, SpectraNet contributes a valuable tool for building more sustainable and cost-effective cloud infrastructures.
Optimizing CAM Refinement with Swin-based Feature Affinity for Weakly Supervised Semantic Segmentation
(Department of Computer Science and Engineering(CSE), Islamic University of Technology(IUT), Board Bazar, Gazipur-1704, Bangladesh, 2025-10-25) Farzana, Anika; Ahsan, K. M.Abesh; Amin, Sayemah
Semantic segmentation is a fundamental computer vision task requiring pixel-level understanding of images. While fully supervised methods achieve high accuracy, they rely on costly pixel-level annotations. Weakly Supervised Semantic Segmentation (WSSS) mitigates this by using weaker supervision, such as image-level labels, to train effective models. Recent WSSS progress leverages Class Activation Maps (CAMs), though their sparsity and poor boundary localization remain challenges. This study enhances CAM quality through multi-modal backbones like UniCL and hierarchical transformers such as Swin Transformer for stronger feature extraction, coupled with an affinity-based framework that fuses encoder and decoder affinities for semantically coherent pseudo-labels. A Pixel-Adaptive Refinement (PAR) module further improves object boundaries using local similarity cues. Experiments on the PASCAL VOC 2012 dataset yield mean IoUs of 50.3% (validation) and 50.8% (test), with strong performance on large, distinctive classes but weaker results for small or human-centric ones due to CAM bias and dataset imbalance. Overall, our findings demonstrate that UniCL and Swin Transformer significantly improve CAM quality and segmentation under weak supervision while highlighting the need for strategies that handle object size variation and reduce model bias.