Thesis (Bachelor of Science in Computer Science)

Browse

Has files: NoSubject: search.filters.subject.Deep learning

Search Results

Now showing 1 - 4 of 4
  • 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
    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
    Ejection fraction estimation using deep semantic segmentation neural network on 2D Echocardiography data
    (BRAC University, 2020-04) Khan, Abde Musavvir; Shejuty, Myesha Farid; Talukder, MD. Nafis Shariar; Zubayear, Syed Ibna; Alam, Md. Golam Rabiul
    Ejection fraction value denotes how much blood is pumped out of the heart to different parts of the body. It is a routine clinical procedure in heart function assessment, where the left ventricle of the heart has to be manually outlined by doctors in clinical settings to measure the EF value which is time consuming and highly varies by observer. Modern day deep learning methods are able to automatically complete this type of outlining task automatically with much ease and better efficiency even when the model is trained on a deeper neural network and smaller dataset. This paper investigates the deep semantic segmentation networks to find the most accurate one to implement an EF estimation system could be built on the most accurate image segmentation network which will reduce the pressure off the doctors shoulders and stop the eyeball estimation of EF values which is subject to inter-observer variability. This paper evaluated three different image segmentation neural networks namely U-Net, ResUNet, Deep ResUNet to find their accuracy score basing mostly on the dice accuracy metric. The most accurate model of the three Deep ResUNet has been utilized to form Left Ventricle segmentation network for end systole and end diastole images on which volume measurement formula is applied to find out the Ejection Fraction value.