Projects

Cycle Thief Detection from Realtime Footage using YOLOv5 and DeepSORT [Project] [Sample]

Ashraf Ul Alam, Soumit Das

This project involves the development of a real-time cycle thief detection system utilizing advanced technologies. YOLOv5 is employed for object detection, enabling the system to identify cycles in realtime footage. DeepSORT is integrated for tracking the detected objects across frames, while the KD-Tree algorithm enhances the efficiency of nearest neighbor searches. Additionally, the Face_Matcher tool is used for facial recognition, further aiding in the identification of suspects. Together, these components create a robust solution for monitoring and addressing cycle theft in real-time.

Cyber Threat Report Summarization Using FLAN-T5 with LoRA Adaptation [Project] [Dataset]

Ashraf Ul Alam

This project focuses on improving the summarization of cyber threat reports using the FLAN-T5 model with Low-Rank Adaptation (LoRA). FLAN-T5 is a pre-trained transformer model, optimized for instruction-based tasks, making it ideal for generating concise and accurate summaries. LoRA is integrated to efficiently fine-tune the model by reducing the number of trainable parameters, thus enabling performance improvements even on limited hardware. The approach leverages structured cyber threat reports to train the model, allowing for quick decision-making in threat mitigation. The effectiveness of the model is evaluated using various NLP metrics, including ROUGE and BERTScore, demonstrating superior performance over the base model.

KD-UDA: Knowledge Distillation-based Unsupervised Domain Adaptation for Improved Medical Image Segmentation [Thesis]

Ashraf Ul Alam, S. M. Mahedy Hasan

The KD-UDA project focuses on enhancing segmentation performance across unseen target domains using knowledge distillation for unsupervised domain adaptation. The primary goal is to adapt models to new domains without requiring labeled data from these domains. This is achieved by transferring knowledge from a model trained on the source domain to a target domain model using knowledge distillation. The approach incorporates source loss and domain shift loss with Kullback-Leibler (KL) divergence to address domain shift issues. The framework’s effectiveness is evaluated on 2D datasets (Drishti-GS, RIM-ONE-R3, REFUGE Source-1, and REFUGE Source-2) for retinal images and a 3D dataset (BraTS2021) for MRI scans. In both scenarios, the framework significantly improves segmentation performance on target datasets.

Parameter Efficient Fine-tuning of DistilBERT with LoRA for Phishing URL Detection [Project]

Ashraf Ul Alam

This project focuses on detecting phishing URLs using the DistilBERT model, enhanced with Low-Rank Adaptation (LoRA) for improved performance. Phishing URLs, which deceive users into visiting harmful websites, pose significant cybersecurity risks. The model classifies URLs as either Phishing or Non-phishing using a sequence classification approach. The dataset, Phishing URLs, is used to train and evaluate the model. LoRA is applied to adapt the pre-trained DistilBERT Base Uncased model, allowing for efficient training and fine-tuning while reducing the number of trainable parameters. This adaptation results in significant improvements, achieving an F1 score of 84.23% after fine-tuning, compared to 57.88% with the base model. The project highlights the effectiveness of LoRA in enhancing phishing detection accuracy without overfitting.

Fake News Detection Using NLP: A Study on BERT and LSTM with GloVe Embeddings [Project]

Ashraf Ul Alam

This project explores two distinct approaches to fake news detection. The first model employs GloVe embeddings with an LSTM-based neural network, incorporating extensive text preprocessing techniques to extract key features such as word frequency, punctuation usage, and stopwords. The second model utilizes BERT, a transformer-based language model, to analyze news titles with contextual word embeddings for improved accuracy. By comparing these methodologies, the study evaluates their effectiveness in detecting misinformation and highlights the impact of deep learning architectures in NLP-based fake news classification.

NeuroSeg3D: 3D Attention U-Net for Accurate Brain Tumor Segmentation (BraTS 2021) [Project]

Ashraf Ul Alam

NeuroSeg3D is an advanced 3D U-Net architecture enhanced with residual blocks and spatial attention modules, designed to effectively capture fine spatial features from MRI images. Utilizing the BraTS 2021 dataset from Kaggle, the model aims to accurately segment brain tumors by focusing on relevant spatial information. The training process was stable, with smooth convergence, demonstrating effective learning and minimal signs of overfitting. Quantitative results show that NeuroSeg3D achieved a Mean Dice score of 84.42% and a Mean IoU score of 75.86% on the validation set, highlighting its strong performance and generalization capabilities in differentiating tumor regions from healthy tissues in MRI scans.

Optimizing Feature Representation of Deep Neural Networks for Enhanced Deepfake Detection [Project] [Poster]

Ashraf Ul Alam, Sudipta Progga Islam

This project focuses on detecting deepfake images using the 140k Real and Fake Faces dataset from Kaggle. The VGG16 model, known for its deep architecture, was employed as the primary feature extractor. To enhance performance, a channel attention mechanism was introduced, allowing the model to prioritize relevant feature channels while reducing the impact of less useful ones. This resulted in a significant improvement in classification accuracy. Additionally, an ablation study was conducted using ResNet50, demonstrating how attention mechanisms improve feature representation. The final model achieved a high accuracy of 99.80% with VGG16 and channel attention, making it an effective solution for detecting deepfake images.

Chronic Kidney Disease Prediction using Machine Learning [Project]

Ashraf Ul Alam

In this project, a machine learning model was developed to predict Chronic Kidney Disease (CKD) at an early stage. The process included comprehensive exploratory data analysis and feature engineering to optimize model accuracy. A user-friendly prediction tool was then created, featuring a Flask API for deployment and a web interface designed with HTML and CSS. This tool facilitates early CKD detection, providing users with a practical solution for managing their health.

Maternal and Child Health Care [Project]

Ashraf Ul Alam, Sudipta Progga Islam

This initiative involved creating a comprehensive Maternal and Child Health Care website designed to support expecting mothers with tools for due date calculation and immunization schedules. The site features personalized SMS and email notifications, as well as a query posting function to facilitate communication with healthcare specialists. Additionally, a mobile app was developed using Android Studio and Firebase, mirroring the website’s features. This project aims to enhance prenatal and postnatal care by providing a supportive online platform for mothers.

Cardiotocogram Data Analysis for Fetal Health Classification Using Machine Learning [Project] [Slide]

Ashraf Ul Alam

This project aims to classify fetal health status using machine learning models applied to Cardiotocogram (CTG) data. The dataset consists of 2,126 records with 21 features extracted from CTG exams, categorized into three classes: Normal, Suspect, and Pathological. Models like Random Forest, K-Nearest Neighbors, and Gradient Boosting were trained on the preprocessed dataset. Feature selection, data standardization, and SMOTE were employed to improve model performance. Random Forest achieved the highest accuracy of 98.47%. This project showcases how machine learning can assist in automating and improving fetal health assessment.

Implementation and Analysis of Neural Networks for Liver Disease Diagnosis [Project]

Ashraf Ul Alam

This study is part of my Neural Networks course, where I implemented and analyzed various neural networks and machine learning algorithms from scratch to gain a deeper understanding of how these models function. The project focuses on classifying liver disease using the Indian Liver Patient Dataset and explores a variety of learning techniques. The study includes the development of models such as K-Nearest Neighbors, Single Layer Perceptron, Multi-Layer Perceptron, Kohonen Self-Organizing Map, and Hopfield Neural Networks. An Exploratory Data Analysis (EDA) was conducted to gain insights into the dataset before model training. Detailed reports and code for each algorithm are provided, showcasing the analysis, findings, and implementations of the algorithms.