Bachelor of Science in Computer Science and Engineering

Browse

Now showing 1 - 20 of 1029

2D Log gabor and SVD based parallel texture feature extraction usingNVIDIA GPU
(BRAC University, 4/20/2016) Raja, Sharif Ahmmad; Ratul, Aminur Rab; Niloy, Sakib Anjum; Uddin, Jia
Texture feature is one of the most popular technique in image segmentation, classification, retrieval and many others. Now a days, among other ways of texture feature extraction, Gabor filtering has been widely used. Here, we are presenting a well ordered two dimensional texture feature extraction method. First, we convert the image to gray level. Then a 2D Log Gabor filter with different frequencies decomposed with the SVD algorithm applies on each converted part of gray level image to extract appropriate distinctive texture information. To evaluate the performance of proposed model, we utilize singular values of SVD as a feature vector. For classifier, we use Naïve Bayes classifier for training and testing our experimental dataset. In our experimental set up we utilize an NVDIA GeForce GTX780 graphics card. Experimental result showed this parallel implementation of our model is56X faster than conventional CPU implementation.
3D fabrication of food through software implementation for patients of various diseases and dysphagia
(BRAC University, 2018-07) Farha, Anika; Muhtadi, Mantaka; Morshed, Ramisa Ibnat; Alam, Md. Ashraful; Islam, Md. Saiful
3D printing can be considered as a means of creating solid 3-dimensional objects using additive manufacturing with the aid of computer design software. With respect to time and newer technologies, 3D printing has advanced far more than we have ever anticipated. Though the potential of 3D printing has been under philosophical discussion for some time, creating food using 3D printers has always been a challenge. In order to print 3D food using 3D printing technology, a solid software is required that will count and personalize the nutrient information per patient and individual. In this research, a software has been developed that computes the required macronutrients according to the requirements per individual. Seven diseases have been outlined in this thesis and calculations for ingredients based on different ratios and equations have been made which have been derived after intensive research, from which two unique self-deduced algorithms have been developed to run the software logic. The system takes in the age, height, weight, gender and disease information from the hospital patients’ charts, uses the variables to run the first own algorithm to calculate personalized protein, fat and carbohydrate counts and then based on the diseases the software runs the second own algorithm to calculate the exact amounts of the ingredients required to prepare their meals per day. In this thesis, 3D models have also been designed to maximize user experience and the patient’s visual appeal for the food. The final output of the software is expected to print out a list of the ingredient count per day and display the selected 3D model. Further work of this thesis aims to explore the world of 3D printing and come up with nutrition specific solutions for people suffering from different diseases and have 3D printing technology make their lives easier because the potential of 3D printing of food is vast in a sense that it could provide an exciting alternative to help people customize their food, make them visually more appealing so that people requiring special care would be able to consume food with much more ease which would otherwise have been unachievable through conventional cooking and food preparation techniques
3D model generation of real-objects using depth and color information
(BRAC University, 2018-04) Sakib, Alium Basir; Azam, Hasnat; Jui, Tonni Das; Akter, Fahima; Alam, Md. Ashraful
We propose and demonstrate a method of generating 3D model using depth and color information of real-object using Microsoft Kinect sensor using our proposed pixel mapping algorithm. The primary idea of the method is to manipulate RGB and depth data within arrays to plot in three dimensional frame to generate the 3D colored model of any real-object. The proposed system comprises of 3 phases. The phases include acquisition of color and depth information of real objects, extraction and processing of RGB and depth data, generating 3D model using our proposed pixel mapping algorithm. Whereas, our proposed algorithm focuses on creating several 1D arrays for depth as well as the color data meaning Red, Green, Blue pixel values separately. By mapping the depth values accordingly pixel-wise 3D black and white model can be created where the background is also included. By limiting the depth values according to the objects position the background can also be removed to construct the focused object's 3d model. Furthermore, our algorithm maps the R, G, B arrays in the black and white model to add proper colors pixel-wise .As a result, the colorful RGB 3D model can be achieved.
3D Visualization of 2D/360° image and navigation in virtual reality through motion processing via smart phone sensors
(BRAC University, 2018) Aurini, Maliha Tasnim; Islam, Shitab Mushfiq-ul; Alam, Dr. Md. Ashraful
The 360° images or regular 2D images look appealing in Virtual Reality yet they fail to represent depth and how the depth can be used to give an experience to the user from two dimensional images. We proposed an approach for creating stereogram from computer generated depth map using approximation algorithm and later use these stereo pairs for giving a complete experience on VR along with forward and backward navigation using mobile sensors. Firstly the image is being segmented into two images from which we generated our disparity map and afterwards generate the depth image from it. After the creation of the depth image, stereo pair which is the left and right image for the eyes were created. Acquired image from the previous process then handled by Cardboard SDK for VR support used in the Android devices using Google Cardboard headset. With the VR image in the stereoscopic device, we use the accelerometer sensor of the device to determine the movement of the device while head mounted. Unlike the other VR navigation systems offered (HTC Vibe, Oculus) using external sensors, our approach is to use the built-in sensors for motion processing. Using the accelerometer reading from the movement, the user will be able to move around virtually in the constructed image. The results of this experiment are the visual changes of the image displayed in VR according to the viewer’s physical movement.
A Blind assistance and navigation system using SIFT algorithm for indoor environments
(BRAC University, 2013-12) Zereen, Aniqua Nusrat; Kabir, Sadia Binte; Alom, Md. Zahangir
When blind people are trying to navigate the city streets, they can get assistance from a speaking GPS-enabled Smartphone, just like everyone else. Once they move indoor and lose access to the required satellite signals. While there are some indoor navigation systems that require things like radio-frequency tags to be strategically placed around the building, it‟s currently unrealistic to expect to find such systems installed in many places. However, we are going to propose blind assistance and navigation system for indoor environment. It required a computer with camera loaded up with digital images of the room. Features and key points of the images have been extracted using Scale Invariant Feature Transform (SIFT)[1] algorithm from parent and child images. Database has been created with the key feature points and sound files for individual child images. Moreover, our Matching algorithm is modified K-D tree[6] known as Best Bin First (BBF) method. It has been used to match the key points of parent images with database to recognize the desire child images. System has been developed to determine the desired object and guide them along it via verbal cues from the database. Experimental results are demonstrated this proposed method is accurate and reliable to solve the promising and challenging Blind Assistance and Navigation System using SIFT Algorithm for Indoor Environments.
A blockchain and capability based access control for internet of things using self-sovereign identity
(BRAC University, 2024-12) Amin, Md. Safinur; Khan, Abu Bakar Siddique; Ahmed, Istihad; Sajid, Tawsif Mahamud; Monir, Md. Mohtasim; Ferdous, Md Sadek
Access control systems are essential tools for businesses because they guarantee the safe management of user access to resources after verification. Conventional approaches, such as Role-Based Access Control (RBAC) and Discretionary Access Control (DAC), have drawbacks such complexity brought on by "role explosion" and susceptibility to Trojan horse assaults. Moreover Attribute-Based Access Control (ABAC) has scalability challenges. As the number of attributes increases, managing and maintaining the policies associated with these attributes can become complex. Another significant drawback is, ABAC relies heavily on policies that define the rules for granting or denying access based on attributes, introducing the challenge of policy management complexity. While it provides an alternative, Attribute-Based Access Control (ABAC) has drawbacks such as “attribute explosion" when the number of attributes rises. The lightweight and dynamic properties of Internet of Things (IoT) devices provide security challenges for centralized access control systems. To ensure that only authorized entities may connect with certain IoT resources, access control becomes essential to handle issues related to single-point failure, user authentication, and privacy leaks. To mitigate the risks of centralized approaches, a technique called distributed access control is suggested. It is investigated whether integrating blockchain technology might improve security. Benefits of blockchain include its decentralized, transparent, and unchangeable nature. Nevertheless, there are drawbacks to the current blockchain-based IoT access control solutions, such as their vulnerability to Distributed Denial of Service (DDoS) assaults. To solve these problems, the suggested method integrates capability-based access management with Self-Sovereign Identity (SSI) inside a blockchain environment. By ensuring that entities only have permissions that are in line with their roles and responsibilities, this method seeks to securely manage and transfer capabilities. The creation of a prototype system highlights the feasibility and effectiveness of the suggested solution in the research. This system provides a possible remedy for the security issues raised by IoT settings and blockchain technology by demonstrating the combination of SSI and capability-based access control in a practical setting.
A comparative analysis of application-level and system-level container runtimes for state-of-the-art data deduplication techniques
(BRAC University, 2025-10) Sarwar, Al-Nahian Bin; Chowdhury, Rohit; Mahin, Sadman Taufiq; Akib, Mohammed; Mukta, Jannatun Noor
Containerization has become a cornerstone of modern software deployment, offering lightweight isolation and rapid scalability across diverse environments. However, the growing variety of container runtimes introduces uncertainty regarding their behavior under data-intensive workloads such as deduplication, where computational efficiency and resource utilization directly affect scalability and responsiveness. To investigate this, we design a structured experimental pipeline that executes three hash-based deduplication algorithms - CRC32, MD5, and SHA-256; within three container runtimes: Docker, LXC, and Podman. Each algorithm is run ten times across datasets of 1M, 5M, and 10M records to ensure statistical consistency, generating over 3,700 performance samples consolidated into 180+ representative instances. Building on this pipeline, we develop a holistic scalability assessment framework that quantifies container efficiency through throughput trends, variability in CPU and memory usage, and collision rates, offering a comprehensive perspective on runtime behavior. Experimental findings show that Docker maintains balanced scalability with stable throughput growth through efficient daemon-managed scheduling, while LXC delivers superior computational efficiency under heavy workloads due to its direct kernel namespace access. Podman, though optimized for lightweight and security-focused tasks, demonstrates performance variability when scaled. Finally, we introduced a decision tree to assist in selecting optimal container–algorithm configurations tailored to workload requirements. This work establishes an empirical foundation for understanding container performance in deduplication contexts, providing actionable insights for building efficient and resilient cloud-native data processing infrastructures.
A comparative analysis of deep learning and hybrid models to diagnose multi-class skin cancer
(BRAC University, 2023-05) Nawrin, Ishrat Nur; Trina, Tonusree Talukder; Rasel, Annajiat Alim; Rahman, Rafeed
Skin cancer is one of the most lethal and increasingly prevalent cancers in the world. Skin cancer develops when the epidermal (top layer of skin) cells divide abnormally, causing it to spread to other regions of the human body. Skin cancer exists in seven different varieties. The presence of malignant epidermal cells determines the type of skin cancer. Dermoscopy, spectroscopy, and imaging tests are primarily utilized to identify the malignancy. These procedures are expensive and prolonged. It may result in unfavorable effects such as bleeding, bruising, and infection as well. The narrow variances in multi class cancer pictures escalate the complexity of classification. Dermatologists confront challenges in the categorization of cancer types from images. Deep learning has resulted in a dramatic leap in disease identification. Deep learning models are capable of categorizing skin cancer more precisely than dermatologists. Several studies focused on pre-trained and hybrid models for categorizing the classes of skin cancer. In contrast to binary classification, the multi-class classification of skin cancer yielded an insignificant result for both deep learning and dermatologists. The proposed study employs varieties of deep learning and hybrid models to examine the performance of each model in categorizing the classes of cancer. The proposed CNN-SVM-LSTM hybrid model obtained the highest result compared to other models, with 87.15% accuracy, 87.42% precision, 87% recall, and 87.428% F1 score. To illustrate the overall comparison of the models, each model has been depicted through a classification report and a confusion matrix.
A comparative malware analysis of XWorm and Nanocore: laying the groundwork for enhanced detection strategies
(BRAC University, 2024) Shanto, Shakil Islam; Punom, Afsana Mimi; Haque, S M Taiabul; Ahmed, Md. Faisal
Malware continues to evolve, posing a significant challenge to global cybersecurity through sophisticated techniques such as obfuscation, process injection, and persistent Command-and-Control (C2) communication. This study conducts a comparative analysis of two active malware strains, XWorm and NanoCore, to uncover shared tactics and unique features that enable their evasion and impact. By employing static, dynamic, and reverse engineering analyses, the research identifies commonalities in delivery methods, persistence mechanisms, and payload obfuscation. Tailored YARA rules are developed to enhance malware detection frameworks, providing practical tools for real-world applications. The findings emphasize the importance of behavior-driven detection strategies and propose generalized mechanisms to address broader malware families, moving beyond isolated strain analysis. This research not only bridges critical gaps in understanding modern malware but also lays a foundation for scalable and adaptive defense systems, contributing to the ongoing battle against evolving cyber threats.
A comparative study of car image generation quality using DCGAN and VSGAN
(BRAC University, 2022-05) Shayer, Mirza Ahmad; Anjum, Nafisha; Mim, Sushana Islam; Chowdhury, Md. Abu Sajid; Preoshi, Noshin Nanjiba Islam; Mostakim, Moin
In today’s modern society, image generation (synthesis) has a great number of uses in various tasks. Image generation is used in crime forensics, improving image quality and generating better images. In 2014, a scientific breakthrough occurred in the machine learning community when Ian Goodfellow and his colleagues introduced the GAN (Generative Adversarial Network). Ever since then, GANs have become a more popular concept in the scientific community. Even today, GANs are being used, utilized and upgraded. This thesis is a comparative study of two GANs used for generating images of cars- DC-GAN (Deep Convolution) and VS-GAN (Vehicle Synthesis). The study will determine which of the two is better suited to generate high quality images of cars. We will train both GANs using the same dataset. The dataset consists of about 16185 Google images of random cars, 8144 for training and another 8041 for testing. The dataset is already preprocessed and split. We will compare the GANs training times, losses, accuracies and pictures generated, showing how well they perform. We will run all the GANs for 40 epochs in both training and testing. We will compare the CGAN, DCGAN, VSGAN, WGAN and WGAN-GP, to see which performs the best. We have used K-Nearest Neighbors, Regression and Random Forest Classifier to calculate the accuracies of all the GANs. We have displayed the results in tabular and graphical formats. We believe this will improve GAN research by providing an excellent comparison between the GANs and determine which is better suited for the given task. We also hope to improve the models further in the future and make an even more in depth comparison between the GAN architectures.
A comparative study of Gemini generated personas and real user personas for restaurant POS design
(BRAC University, 2026-02) Ahmed, Junaed; Rabiul Alam, Md. Golam
The work experience under the internship is recorded in this report, which is based on the position of a UI/UX Designer in the design of a Restaurant Point of Sales(POS) Design at Bdcalling IT Ltd. In particular, it demonstrates a case study of incorporating Human Computer Interaction (HCI) which focuses more on real world human users prior to the creation of interactive systems, but with the recent increase in the accessibility of large language models (LLMs) like Google Gemini, designers of user personas are evolving their views of the task. The paper presents a personal post internship research that presents a comparative study systematically of personas developed using real restaurant employees versus personas developed using Gemini of similar positions in a restaurant point of sale (POS) design. Based on HCI and UX literature on personas and synthetic users, the study develops and evaluates the hypothesis that Gemini generated personas can be used to approximate real user personas regarding the high level attitude and task objectives but it is systematically under representative of local infrastructural, organizational and cultural constraints. A structured questionnaire was used to gather new field data with the restaurant managers, cashiers, waiters and kitchen staff, and similar Gemini personas were generated by using detailed prompts to describe the same scenario. The most significant constructs included learning comfort with new software (1-5) and the rating of the importance of POS features (1-3) that were aligned among the groups and compared with the Mann Whitney U test that was adequate in the case of small and independent samples with ordinal Likert type data. Findings indicate that Gemini personas are quite close to actual staff on such large constructs as comfort with technology and the value of fast order entry and split bill support, and are more conservative in terms of the strength of issues relating to stock visibility, offline resilience and staff level accountability as reported by workers. The paper concludes that Gemini generated personas can be useful to jumpstart early HCI work, particularly when resources are limited, but it must be applied as a baseline and confirmed against actual user data before it is applied to make critical design choices in human centered software engineering.
A comparative study of machine learning and geospatial techniques for analyzing Dengue diffusion patterns and identifying hotspots in Bangladesh
(BRAC University, 2025-01) Siddika, Taskia; Ethuna, Shuria Akter; Progga, Nafisa Ahmed; Ratul, Niamotullah; Kamal, Mirza Fahad Bin; Alam, Md. Golam Rabiul; Rahman, Rafeed
Dengue fever is still a major public health challenge in tropical and subtropical coun- tries, especially in Bangladesh where epidemic has been a big threat to public health. In this paper, we have developed an integrative computational approach analyzing geographic information and employing data mining to forecast dengue spread and reveal vulnerable regions. Our reference methods include Ordinary Least Squares (OLS), Geographically Weighted Regression (GWR), Lasso Regression, Elastic Net Regression, Bidirectional Long Short Term Memory (BiLSTM), and DiffFlow as well as TabDDPM. The study builds on past results, climatic parameters, and population density to improve predictive performance. The data set used in this research was collected from the official web site of Directorate General of Health Services (DGHS) that made the data authentic. The proposed approach, as a result, provides sig- nificantly higher predictive performance than conventional statistical analysis based on the spatial and machine learning components. Furthermore, to categorize and prioritize the high-risk areas, we apply special methods of multiple criteria decision making – TOPSIS and VIKOR. We reveal that the proposed models based on ma- chine learning methodologies are useful for identifying dengue fever hotspot areas and enlightening information for public health officials regarding timely application of control measures.. This study emphasises the necessity of the epidemiological and climate data coupled with the computational modeling to mitigate future outbreaks."
A comparison between open-source and commercial translator tools
(BRAC University, 2007-01) Rahman, Atiqur; Abdullah, Matin Saad
When staring to analyze the tools used for translation, I have chosen Pootle as the first tool to consider is an open source tool that performs translation by means of analyzing the whole sentence first before parsing for translation. This tool is used by Ubuntu, leading distribution of Linux, and because it is open source, it can easily be modified to one’s own language and hence be extended. Then for studying a commercial translation tool, the popular tool that I have found and studied is Trados. Despite being closed –source it has a good tutorial and documentation for use. It works by creating a project and keeping the file to translate then translating it. However, the creation of project is only available to the software that is bought with a certain price. Performance wise, Trados has a different way for translating, it translates each word before going for grammatical parsing. This might give rise to certain difficulties or even inconsistencies in the translation process and eventually lead to grammatically complex or incorrect sentences. However, Trados has ways of correcting for these issues but doesn’t document how they do it as they are closed-source. Now, which method to use is a debate that needs to be resolved in a scientific way. Hence this project is concerned with the analysis between the two leading yet distinct tools for translation.
A comprehensive NLP-based voice assistant system for streamlined information retrieval in metro rail services of Bangladesh
(BRAC University, 2024-01) Rahman, MD. Abir; Alamgir, Adnan; Chowdhury, Shaheedul Haque; Mushtari, Maliha; Anzum, Wasim; Rahman, Md. Khalilur; Artificial intelligence
Bangladesh’s capital city Dhaka is served by the Dhaka Metro Rail. A metro-rail based rapid transit system is considered one of the important technologies that may decrease the working-hour wasting issue in a developed nation owing to traffic congestion. It moves between communities within an urban region or the cities that constitute a metropolitan area. However, to reap the advantages for both passengers and metro-rail authorities, a voice assistance system is also necessary for a metro-rail-based transit system. Many passengers expressed their dissatisfaction and frustration at the appearance of such difficulties from the very beginning of the metro-rail service. Many people complained about experiencing trouble obtaining tickets from vending machines due to technological challenges when the mass transit system was opened to the public. Officials reported that vending machines stopped operating as people attempted to use them without understanding how to use them. This research proposes a noble approach for the general population of Bangladesh. General people will be able to interact with a voice assistant and get their job done, such as collecting information about the train and metro-rail station. We will be undertaking our research with the help of Natural Language Processing (NLP) based on the Artificial Intelligence Markup Language (AIML) structure for training the model. The primary dataset creation procedure is cautiously defined, comprising question generation, response formulation, and category assignment. To ensure the relevance and accuracy of our dataset, a thorough verification procedure was done in collaboration with the Managing Director of Dhaka Mass Transit Company Limited (DMTCL). Term Frequency-Inverse Document Frequency (TF-IDF), and a sequential neural network model are trained with the dataset. We designed a web application with the capability to receive voice input and provide spoken output. This application was developed by utilizing a voice recognition Application Programming Interface (API) for voice-to-text and text-to-voice conversion. A closed domain question answering (cdQA) NLP solution was utilized to acquire information about the given query. The paper intends to show how voice assistants can be used in daily life in metro rail stations with minimal effort and to analyze if there is potential for making them accessible to the general public.
A comprehensive safety and support platform for domestic abuse victims
(BRAC University, 2025-06) Mahin, Rifah Tasnim; Ainun, Atika Hossain; Islam, Lamiya; Chowdhury, Farida; Abedin, Jawaril Munshad
Domestic violence remains a critical issue, especially in surveillance-heavy environments like Bangladesh where abusers often monitor their victims’ mobile activity. This research presents the iterative design and conceptual development of a discreet safety and support application for domestic abuse victims. Unlike traditional solutions, this mobile application is disguised as a benign utility app (e.g., a grocery list), ensuring discretion even under close monitoring. The app architecture follows a layered Four P’s Model: Preparation, Protection, Provision, and Prevention, aligning each feature with user safety goals. While designing, a user-centered approach was adopted, involving expert interviews, focus group discussions, feature assessment surveys, and victim testing across three design phases: hand-drawn paper prototypes, low-fidelity digital versions, and a fully navigable high-fidelity prototype. Each phase incorporated active feedback from survivors and professionals to ensure clarity, minimal cognitive load, and cultural relevance. Key functionalities include dummy interface switching, real/dummy login system, encrypted evidence logging, a Quick Exit button, and Bangla localization. Additionally, the app proposes two machine learning extensions: a voice-based distress and trigger word detection model using emotion recognition, and a conceptual risk prediction framework based on user-logged incidents. While not implemented due to time and development limitations, the models were architected using open-source datasets and preprocessing pipelines, ensuring future feasibility. By embedding iterative victim feedback and Human-Computer Interaction (HCI) principles throughout, this study demonstrates a survivor-informed, context-sensitive approach to mobile safety design. The final prototype serves as both a practical intervention model and a contribution to ongoing research in HCI, trauma-aware design, and machine learning for social good.
A comprehensive study for predicting eyesight disease using ML
(BRAC University, 2024) Sayem, Tanvir Islam; Sara, Fouzia Rahman; Biswas, Poroma; Bhowmick, Debabrata; Rhaman, Khalilur
From mild to severe distant vision impairment caused by untreated conditions, including cataract, glaucoma, retinal disease, and diabetic retinopathy, more than60% of the world’s population—exceeding 4.5 billion individuals—requires corrective lenses or treatments for visual and retinal disorders. The fundamental goal ofthe current study is to create an advanced deep learning (DL) system capable ofcategorizing retinal pictures into five groups. A deep convolutional neural network(CNN) was used to classify normal eyes, cataracts, glaucoma, retinal illness, and diabetic retinopathy. The dataset, obtained from Kaggle, had 2827 pictures that wererandomly divided into training, validation, and testing groups. The TensorFlowobject identification framework was used to create many CNN meta-architectures,including YOLOv5, YOLOv7, and InceptionResNet50. The YOLOv5 model showedgreat development. The YOLOv5 model demonstrated significant progress in detecting the mentioned eye diseases and achieving 0.951 mAP for 7357 images.
A comprehensive study of networking systems and their real-world implementation ‘’from theory to practice: my internship experience in SYSOLUTION”
(BRAC University, 2025-11) Tamanna, Sanjida; Islam, Md. Saiful
SYSSOLUTION is a Bangladesh-based software and IT solutions company offering a range of services across various sectors, including healthcare, real estate, fintech, telecom and e-commerce. This business has established itself as a pioneer in the telecom sector, delivering exceptional customer care and ensuring its presence at every step of the value chain. Moreover, it has a data center of its own to support cloud hosting and its managed services portfolio. Cloud-based solutions, enterprise systems, web and mobile application development, and custom software development are all included in the company’s portfolio. SYSSOLUTION seeks to improve user engagement, operational efficiency, and long-term growth for its clients by utilizing cutting-edge technologies and a client-centric strategy. I have learned how to manage a system and generate new ideas from my work at SYSSOLUTION. They encouraged me to learn more about emerging technologies and progress in my network engineering profession. My networking abilities have increased as a result of working with them. Therefore, I have gained new ideas about networking and customer management by working there.
A conventional & deep learning strategy for analyzing & detecting Bengali fake news in online medium
(BRAC University, 2023-04) Ahmed, Istiak; Prima, Shanzida Binta Akram; Baptee, Tahsin Anzum; Afroz, Mehrin; Shanto, Ariful Islam; Rasel, Annajiat Alim; Sadeque, Farig Yousuf
Nowadays, social networking sites like Facebook and Twitter have become an sig- nificant impact on our lives . We use such sites to remain in touch with one another and as a source of news to stay informed about current events. As a result, we frequently see news articles with click-bait headlines from various web portals that lack authenticity. The majority of these sites that share these sorts of links are used to manipulate people and spread false propaganda. We intended to utilize both traditional machine learning algorithms and deep learning algorithms on manually annotated data-sets to create effective approaches for spotting Bangla fake news on online media that included about 8,500 pieces of news data. In particular, in this project, we used classic machine learning algorithms for text classification such as “Naive Bayes Classifier”, “Support Vector Machines (SVM)”, “K-Nearest Neigh- bor (KNN)” as well as other classification-based algorithms such as “Decision Tree (DT)”, “Logistic Regression (LR)”, “Random Forest” and “AdaBoost”. We have also used deep learning models based on Feed-Forward Neural Networks such as “Convolutional Neural Network (CNN)” as well as a variety of Recurrent Neural Networks (RNN) such as “Long-Short Term Memory (LSTM)”, “Gated Recurrent Unit (GRU)” to detect fake news on online media. To conclude, our research focused on developing precise strategies for spotting fake news on social media sites.
A counseling system to predict the study path for freshmen
(BRAC University, 2019-04) Usha, Rowshni Tasneem; Parvez, Shiny Raisa; Sejuti, Fariha Sazid; Hossain, Maisha; Majumdar, Mahbub
Now a days, dilemma related to one's career has been considered as a serious issue, specially among fresh graduates. Starting at the age of 18, the students usually fail to grasp the idea of which career path to pursue as they lack maturity and experience on the matter. Moreover, students su er greatly in deciding which faculty would result the highest bene t for them due to the insu ciency of counselors in the pre-university education. The students do not have the su cient knowledge to make themselves aware of the real life career related challenges, which is supported by academic majors. It is crucial for a student to make the proper decision on the matter of their career in order to avoid consequences that may be the result of wrong career selection. As a result, selecting an proper career with highest bene t has become one of the most di cult as well as challenging task for the students because wrong career selection may lead to a work eld which was not meant for them. This paper presents a counselling system to predict study path for the freshmen by analyzing necessary attributes such as skills, interests, values and motivation, academic background. Moreover, the proposed freshmen counseling system helps the freshmen in their career choice as well as guides toward their respective appropriate career for future. We have used several di erent approaches for modeling and prediction such as Decision tree classi er, Random Forest, SVM, K-Nearest-Neighbors Classi ers etc. and di erentiated between the resulting precision scores. The results were also cross-checked which determined the best parameters that is responsible for providing highest accuracy scores. Furthermore, some ranking algorithms were used to generate a ranked output for the student counseling system. In this paper, we have separated our work into di erent parts. Chapter 1 contains the overall idea about our work. Chapter 2 contains Related Works, followed by Chapter 3, where we have mentioned about the methodologies which we have used for the system. After that Chapter 4 contains the implementation of the system. Next in Chapter 5 result analysis has been mentioned. Lastly, Chapter 6 ended with conclusion, our limitations and future scopes of improvements related to our work.
A credible, automated e-voting system in the context of Bangladesh
(BRAC University, 2018) Sharif, Fairouz; Al Faiyaz, Abdulla; Arif, Hossain
Elections are believed to be the best possible way to live in democratic era and voting is one of the electoral processes that ensures the alignment of democracy in our society. In this thesis, we propose an electronic voting system which can minimize the flaws of traditional voting system. The manual system of voting includes major chances of vote rigging and manipulation, which can be brought to its barest minimum by our proposed system. Also, in paper-based voting system, great amount of people and money is invested and yet the results are not satisfying. Our proposed voting system promises to reduce these waste and serve as a cost-effective, easy-to-use and secure one. As our country have never fully adopted any electronic voting system before, we have tried to build it in accordance with the steps of procedural voting system. Along with biometric authentication, the system also includes candidate pictures along with their voting symbols, using which voters can easily pick their desired candidate with little amount of time. This system will be effective for both sides, one who are voting and another who are monitoring the whole process.