Thesis (Bachelor of Science in Computer Science)

Browse

Has files: No

Search Results

Now showing 1 - 10 of 17
  • No Thumbnail Available
    Item
    Trustworthy AI-enhanced DevSecOps framework: bridging XAI and RAG for interpretable anomaly detection and context-aware root cause analysis
    (BRAC University, 2026-01) Reza, Nowshin; Parvez, Md. Adnan; Mottakee, Md Sayem; Islam, Ariful; Hossain, Muhammad Iqbal
    Modern software development faces many challenges in balancing rapid delivery with robust security. The challenges are bugs,vulnerabilities and inefficiencies. The AI-based DevSecOps tools that are being used today are effective at detecting abnormalities and vulnerabilities but they act as a black box that provides little to no transparency or context that leaves the security team a little skeptical and makes it hard for them to trust the tools and act on them. An AI-driven DevSecOps system that employs Explainable Artificial Intelligence (XAI), which offers insights into the choices made by the transformer model LogBert, is presented as a solution to this problem. The proposed framework also uses Retrieval-Augmented Generation (RAG) that dynamically retrieves contextual data(i.e. Historical fixes, security protocols and optimization patterns). LogBert is trained to model normal system behaviour, detect anomalies while SHAP boosts interpretability and RAG produces historical context which fosters trust from developers and collaborative debugging to eliminate any doubt which is induced by the black box dilemma. In this framework security is actively strengthened through real-time vulnerability scanning, threat prediction and automated compliance check. Moreover, LLM is combined with SHAP to describe what the output of SHAP means, making it more human interpretable. Also it is applied to RAG which transforms retrieved contextual information into simple human readable texts and proposes fixes. This proposed framework does not only improve performance but it also provides a clear path for using AI in a way that is both sustainable and transparent within DevSecOps test phase.
  • No Thumbnail Available
    Item
    Deep learning-based behavioral drift modeling for continuous biometric authentication
    (BRAC University, 2026-01) Dibya, Suprava Saha; Minar; Abdulla-Al-Noman; Shresto, Md. Zafir Abid; Sayef, Ahmed; Hossain, Muhammad Iqbal
    Ascending technologies like the concepts of robotics, artificial intelligence (AI), the concepts of smart devices and the Internet of Things (IoT) have been assimilated into traditional and indispensable physical, biological and digital systems incubating the fourth industrial revolution. Following the era where the attainability of AI has become ubiquitous, vitality has risen to integrate the smartness of AI into traditional authentication systems in order to strengthen the security and develop a resilient and robust barrier between the infiltrator and the delicate systems. The integration of unimodal physiological biometrics along with the previous password systems served the purpose to a certain extent. But the widespread use of AI has given the intruders an easier access to breach the seclusion of systems, raising a massive question for the security, especially in this era of digital transactions, banking, healthcare systems and academic assessments. There comes the necessity of initiating the utilization of behavioral biometric authentication that integrates the behavioral drift of individuals so that impostors relying on static templates and physiological traits like fingerprints are recognized and the system remains decontaminated. In this regard, this research proposes a two-level keystroke dynamics authentication system that provides defense-in-depth security through continuous behavioral drift monitoring. Level 1 employs fixed-text password authentication using Temporal Convolutional Networks (TCN) with contrastive learning and population-based training (PBT), achieving 3.77% EER for high-security login verification. Upon successful authentication, Level 2 continuously monitors user typing during natural free-text sessions using a 4-Model Deep Learning Ensemble. If suspicious behavioral drift is detected during monitoring, the system re-authenticates the user by returning to Level 1, providing a feedback loop against session hijacking and behavioral anomalies. The system explicitly models long-term behavioral drift through the fixed-text component’s PBT mechanism, which adapts to temporal variations in password typing patterns. This research leverages free-text (Buffalo dataset) and fixed-text (GREYC dataset) keystroke dynamics datasets, extracting temporal and spatial features including hold times, up-down latencies, down-down intervals, and trigram patterns. The final system achieves 96%+ accuracy for fixed-text authentication (3.77% EER) and 83.53% True Acceptance Rate (TAR) at 5% False Acceptance Rate (FAR) for free-text continuous authentication (10.05% EER), demonstrating superior performance and strong resistance to impostor attacks through continuous behavioral monitoring.
  • No Thumbnail Available
    Item
    Classification of Alzheimer’s and Dementia subtypes using R-STDP driven spiking neural networks
    (BRAC University, 2025-10) Fariha, Anika; Tasnim, Noshin Fouzia; Manal, Zafeera; Ira, Rayatun Tehrin; Tanzeem, Eshat; Alam, Md. Golam Rabiul; Humayun, Zayed
    Early classification of Dementia which can further lead to Alzheimer’s disease remains challenging due to subtle brain structural changes in MRI scans. This paper presents a novel neuromorphic feature extraction approach combining biologically inspired temporal encoding with Forward-Forward learning for the classification of Alzheimer’s and dementia subtypes. Our methodology employs a three-stage pipeline: neuromorphic preprocessing converts 32×32 brain MRI data into temporal spike patterns across 20 time steps, incorporating skull stripping and CLAHE enhancement; Forward-Forward learning with Reward-based Spike-Timing-Dependent Plasticity (R-STDP) autoencoder extracts latent features without traditional backpropagation; ensemble classification using Random Forest and Gradient Boosting provides final predictions. The neuromorphic preprocessor generates 8,192- dimensional feature vectors capturing temporal dynamics including first spike timing, burst detection, temporal phases, and activity statistics. The Forward-Forward R-STDP autoencoder learns biologically-plausible representations through positivenegative sample discrimination with a 256-dimensional latent bottleneck. Advanced feature selection reduces combined features to around 3000 optimal dimensions through variance filtering, statistical selection, and recursive feature elimination. Our system achieves 77.92% ensemble accuracy on Demented vs. NonDemented classification, with 79.47% weighted precision and 77.69% out-of-bag score. Random Forest achieves 78.77% accuracy while Gradient Boosting reaches 76.93%. Our neuromorphic approach allows parallel processing, reducing computational overhead compared to conventional deep learning and using biologically-inspired representations to capture temporal patterns in brain imaging data. This framework aims to provide an energy-efficient alternative to traditional deep learning while ensuring that a robust classification performance is maintained for neurodegenerative disease detection.
  • No Thumbnail Available
    Item
    Improving histopathological image classification performance leveraging shared knowledge in a federated learning environment
    (BRAC University, 2025-02) Tahsin, Adiba; Reza, Md Tanzim; Faruk, Farhan
    In today’s data-driven world, safeguarding sensitive data such as health records is es- sential to building a secure digital ecosystem. Federated Learning (FL) addresses these privacy concerns by enabling decentralized model training, where devices perform local training and share only model updates with a central server. This paper investigates the application of FL with shared learning in histopathological image classification using Mo- bileNetV2, VGG19, and DenseNet121 models to enhance convergence speed, accuracy, and scalability across diverse datasets. This research shows that MobileNetV2 achieved the largest reduction on the PatchCamelyon dataset (37 to 13). DenseNet121 outper- formed across all datasets, with reductions from 39 to 18 (PatchCamelyon), 35 to 18 (Breast Histopathology), 85 to 67 (Chaoyang), and 28 to 13 (Histopathology Imagery). VGG19, though less impactful, performed best on the Histopathology Imagery dataset (23 to 9). Furthermore, MobileNetV2 demonstrated a significant accuracy improvement on the highly imbalanced Chaoyang dataset, increasing from 69% to 82%. These findings highlight the real-world advantages of faster convergence and improved efficiency, par- ticularly in resource-constrained fields like healthcare. By leveraging shared knowledge across datasets, our approach enhances model generalization and robustness without cen- tralized data collection, paving the way for scalable, cost-effective, and privacy-preserving AI systems for medical image classification.
  • No Thumbnail Available
    Item
    Attention-deficit/hyperactivity disorder detection leveraging an ensemble of encoder-decoder transformer and XGBoost models
    (BRAC University, 2024-10) Sarker, Sharon Rose; Mehjabin, Saowmi; Piper, Meherin Majid; Rahman, Rafeed
    "Early detection of neurodevelopmental disorders such as Attention-Deficit/ Hyperactivity Disorder(ADHD), can lead to improved outcomes and prompt intervention. Traditional detection methods have been facing challenges due to judgment and misinterpretations, lack of resources, and biasness which may cause under-diagnosing or over-diagnosing. Early detection of these neurodevelopmental disorders, not only helps individuals to get proper ministrations but also it can improve their social, cognitive and mental development. In this study, our aim is to build an ensemble model leveraging a custom Transformer with various attention mechanisms alongside an XGBoost model to improve diagnostic accuracy. By comparing the proposed model with other traditional machine learning and deep learning models, this study aims to enhance the accuracy and efficiency of diagnosis. By using a pre-processed EEG dataset and customized ensemble model, the proposed model has achieved 83% of accuracy, highest accuracy among the traditional models. Moreover, this research aims for future development in the field, by offering methodologies that can be useful to further studies focused on disorder detection. In conclusion, this research will use an ensemble model leveraging a custom Transformer with various attention mechanisms alongside an XGBoost model for early ADHD aiming to create a new precision and accessibility in identifying neurodevelopmental disorders. "
  • No Thumbnail Available
    Item
    SSI-Federation: facilitating identity federation using self sovereign identity for Web-services
    (BRAC University, 2024-06) Mahir, Sajid Imam; Ahsan, Navid Alvi; Eshan, Md. Mehrab Hasan; Ferdous,Md Sadek
    Identity federation means entrusting an entity’s online identity verification to an external organization. Identity Federation’s basic concept is that an IdP or Identity provider ensures an entity’s identity to the SP or the Service Provider an entity that provides web service. This is an old concept having the issue of how securely the information will be gathered and stored. To provide security of personal information and to get an overall convenience efficiently Self-sovereign identity or SSI is used. SSI is different from any other verification system due to its peer-to-peer decentralized system with the help of blockchain. This process provides an entity full control of how much personal information they are sharing and who they are sharing it with, with the convenience of service access without login credentials. This reduces the dependency on a specific third party making the process more secure whilst ensuring proper privacy over their data. In SSI like the Identity Federation, there are also two entities other than the user which are Issuers and Verifiers where issuers are trusted credential providers, and the Verifiers are trusted to verify them when requested. Still, the issue here is that there is no connection between the Issuer and the Verifier which concerns the issue of trust among these two entities. We provide a solution to both of these problems by first using SSI as the base model and then enabling the Issuer and Verifier of it to establish trust among themselves before the user requests a service through SSI. For this to succeed the Verifier will also play the role of the SP and the Issuer can be thought of as the IdP. This hybrid system of ours contains an external trust layer over SSI which makes it function like Federated Identity by also keeping the characteristics of SSI with the help of hyperledger-based blockchain technologies.
  • No Thumbnail Available
    Item
    Pothole detection using lightweight network models
    (BRAC University, 2024-01) Abdullah, S. M.; Hasan, Shakib Al; Parsa, Antara Firoz; Kabbya, MD. Asif Shahidullah; Talukder, Anika Hasan; Noor, Jannatun
    Potholes are defective cavities found on road surfaces. Potholes can lead to serious accidents and vehicle damage if not properly detected. Thus, we are proposing the use of neural network models for pothole classification. The study involves a comprehensive performance analysis of existing lightweight neural network models in pothole classification, compared against the traditional heavyweight models. Lightweight models are emphasized in the thesis due to their low computational requirements, faster prediction times and better compatibility with real-time detection. We have tested six lightweight models (CCT, CNN, INN, Swin Transformer, EANet and ConvMixer) and four heavyweight models (VGG16. ResNet50, DenseNet201 and Xception). A custom dataset of 900 images containing image samples from roads of Dhaka and Bogura was created by the authors to run the models. The dataset was further augmented into 10,000 images by applying various augmentation methods. Separate tests for each model were conducted in the augmented dataset to compare performance against the original dataset. Augmentation enhanced the performance of 9 out of the 10 models. CNN achieved the highest accuracy of 96.55% and the highest F1 score of 0.96 in our testing. Furthermore, CCT exhibited accuracy of 94.6% and F1 score of 0.9. The lightweight models overall performed better than the heavyweight models in both datasets.
  • No Thumbnail Available
    Item
    Content based image search in openstack swift
    (BRAC University, 2021-09) Ali Uday, Mir Rownak; Islam Sakif, Md. Sadiqul; Mukta, Jannatun Noor
    The OpenStack Object Store, also known as Swift, is a cloud storage software. Swift is optimized for durability, availability; also concurrency across the entire data set. However, Swift does not have a proper technique to let users and administrators search inside the object storage without the entire OpenStack Infrastructure. In this paper, we propose a Content-Based Image Model for Swift, which enables us to extract additional information from images and store it into an elasticsearch database which helps us to search for our desired data based on its contents. This novel approach works in 2 parallel stages. First, the image which is being uploaded is sent to our trained model for object detection. Secondly, this information is being sent to the elasticsearch, which in return helps us to do the searching based on the contents of the uploaded images. As the accuracy of the search solely depends on the accuracy of the object detection model, we have trained our model with MS COCO Dataset. Lastly, we upload these images in various segments to find out the efficacy of our model not only in real-life small and medium-size Swift object storages but also as a user-centered Content-based image retrieval system from a text-based database.
  • No Thumbnail Available
    Item
    Semantic segmentation with attention dense U-net for lung extraction from X-ray images
    (BRAC University, 2023-03) Auvy, Akib Al Mahmud; Sharif, Shezhan; Chowdhury, Aseer Iqtider; Elahi, Mahbub-E; Mahmud, Washik Al; Noor, Jannatun
    "In the diverse field of computer science, deep learning and digital image processing plays a vital role in medical image research. With a deep knowledge on hand, we can make a machine understand any medical documentation, and fourth, image segmentation, classification, detection, identification, and segmentation have become more reliable and precious by the day. Lung disease detection is one of the most challenging parts of automation machine detection; to achieve that, segmentation is vigorous. For our research purpose, we aim to seek a better-unused model for lung segmentation, and it is fruitless to justify all the deep learning model as the number is huge. Most of them has already been evaluated by another researcher. This is why we have used U-net architecture (Attention Dense U-Net, Dense U- Net, Attention U-Net, U-Net, U-Net++) to segment the lung from an X-ray image. For the named architecture, we have used two convolutional layers. Four types of accuracy measurement matrices were used to judge this U-Net model: accuracy, Dice coefficient, intersection over union(IoU), and validation loss. The milestone for our research is as follows: our collected dataset was originally 512 x 512 pixels which we converted to 256 x 256 pixels for a 2 x 2 patch. This enables the machine to read the image with a better result. The dataset is annotated and masked. After that, we performed deep learning of the U-Net structure for our dataset to train the U-Net model and segment lung pixels from an X-ray image. In this step, we first omitted the background from the image with the help of true positive, true negative, false positive, and false negative values. Finally, we measured our model accuracy by the advanced accuracy measurement algorithm to justify its capability in terms of unknown data. Following these three steps, we have found that Attention Dense U-Net gives the best accuracy for all given parameters, with the result of accuracy: 97.48% Dice coefficient: 94.87% IoU: 93.87%. And the lowest is base U-Net with an Accuracy score: of 96.68%, Dice Coefficient: of 92.1% IoU: of 91.75%. The study reflects that U-Net is unsuitable for the segmentation of lungs from X-ray images. Hence, we have suggested our approach with Attention Dense U-Net for lung segmentation."
  • No Thumbnail Available
    Item
    A deep learning based autonomous electric vehicle on unstructured road conditions
    (BRAC University, 2021-09) Adnan, Ashik; Rahman, G M Mahbubur; Hossain, Md. Mahafuj; Mim, Mahfuza Sultana; Rahman, Md. Khalilur
    Autonomous driving vehicles are too known as driver-less cars which is one of the foremost astounding advances of the twenty-first century, anticipated to be driverless, effective, and crash dodging ideal urban cars of the future. Autonomous cars actually sense the environment, navigate and fulfill human transportation capabilities without any human inclusion. Cameras, radar, lidar, GPS, and navigational pathways help this type of vehicle detect its surroundings[6]. Even when the conditions alter, advanced control systems interpret sensory data to maintain their locations. Autonomous vehicles are on their way to completely replacing the world’s transportation system. To reach this goal automobile industries have begun working in this zone to realize the potential and unravel the challenges as of now. A few companies have also started their trail. It will aid in reducing traffic, reducing pollution, avoiding maximum accidents, saving time, conserving energy, and improving human safety. As a result, with the aim and vision of eradicating these challenges from our country, we are focusing on an independent car that will assist us in saving ourselves from the daily revelations we generally confront on the road. Besides, it is high time we began working in Bangladesh on a driver-less vehicle.