Spring 2018 Final Exam Schedule

 

Final Examination Schedule

PLEASE JOIN US AS THE FOLLOWING CANDIDATES PRESENT THEIR CULMINATING WORK.

Spring 2018
 

Monday, May 14

Collin Gordon

Chair: Dr. Munehiro Fukuda
Candidate: Master of Science in Computer Science & Software Engineering
11:00 A.M.; DISC 464
Agent-Based Data Science and Machine Learning

The Multi-Agent Spatial Simulation (MASS) library is a library that implements an agent-based programming paradigm in Java, C++, and CUDA. This library has been used to great effect in the parallelization of a variety of simulations and data analysis programs. Building on this foundation, the Agent-Based Data Science and Machine Learning project is an exploration into the advantages using MASS Java to parallelize computationally complex clustering, classification, and graph algorithms.

This project presents the algorithm designs for agent-based versions of K Means Clustering, K Nearest Neighbor Classification, Triangle Counting in Graphs, and the Traveling Salesman Problem. In addition to the designs of the algorithms, we present an analysis of programmability and performance comparing MASS Java to the widely used MapReduce and Spark paradigms. We also explore the contributions of previous graduate researchers and position the project as a launching point for expanding the use-cases for MASS. 

Wednesday, May 16

Wei Xu

Chair: Dr. Yang Peng
Candidate: Master of Science in Computer Science & Software Engineering
1:15 P.M.; DISC 464
A Cloud-based Software Framework for Shared Usage of Bluetooth Low Energy Devices over the Internet

With the development of Internet of Things, numerous wireless IoT devices have been brought into our daily lives. Of all the available wireless communication technologies to these IoT devices, Bluetooth Low Energy (BLE) has become the most widely adopted one due to its design of ultra-low energy consumption and generic service stack. However, because of the short communication range and exclusive usage method, a BLE-equipped device can only be used by a single user who’s near the device. To fully explore the benefits of the BLE technology and make a BLE-equipped device truly accessible over the Internet as an IoT device, in this project, we implemented a cloud-based software framework that can enable multiple users to interact with various BLE-equipped IoT devices over the Internet. This framework includes an Android App (agent program) that can run on any Android device (gateway), a suite of services hosting in the cloud (server) to bridge users and gateways, and a set of RESTful APIs exposed to users. Given the availability of this framework, the access to BLE devices can be extended from local to the Internet scale without any software or hardware change to the device, and more importantly, shared usage of a BLE device among multiple Internet users is also made available to further extend the usability of BLE-equipped IoT devices. The correctness and performance of the framework has been evaluated through extensive tests.

Albert Yee-Him Ng

Chair: Dr. Dong Si
Candidate: Master of Science in Computer Science & Software Engineering
5:45 P.M.; DISC 464
Detection of Beta-Barrels from Medium Resolution Cryo-Electron Microscopy Density Maps

Cryo-electron microscopy is becoming one of the most popular techniques used to determine the protein structure. Using medium resolution cryo-electron microscopy density maps, the secondary structures within proteins can be seen. These secondary structures, local substructures located within proteins, help form the overall structure and shape of proteins. The automatic detection of these secondary structures directly from density maps would be helpful to many, such as biologists and drug researchers. One such secondary structure is the beta-barrel, a beta-sheet based secondary structure, that is commonly found as in cell membranes and transport proteins. This thesis proposes a novel method combining convolutional neural networks, genetic algorithms, and ray casting to perform automatic detection for beta-barrels from within medium resolution cryo-electron density maps. This approach was tested using both experimentally produced and simulated cryo-electron microscopy density maps.

Friday, May 18

Gousiya Farheen Shaik

Chair: Dr. Min Chen
Candidate: Master of Science in Computer Science & Software Engineering
1:15 P.M.; DISC 464
ViDupe- Duplicate Video Detection as a Service in Cloud 

In this project, we aim to create a web service called ViDupe to detect duplicate video files in a user’s personal cloud drive. The existing stand-alone desktop applications find duplicates based on metadata and/or content comparison techniques. All such applications are computationally intensive and are more focused on duplicate image detection. Duplicate video detection is more challenging in terms of long processing time, large size and number of videos to be compared.  

To address this challenge, we propose to improve the algorithm efficiency by enhancing and revising the existing metadata and content comparison methods and enable public access by developing a web service. Currently, the scope of this project is limited to detecting duplicate videos in a user’s personal cloud drive. This approach may be further extended and applied to other cloud storage. 

Craig Shih

Chair: Dr. Munehiro Fukuda
Candidate: Master of Science in Computer Science & Software Engineering
3:30 P.M.; DISC 464
Benchmarking and Evaluating ABM Parallel-Programming Features in Social, Behavioral and Economic Sciences

Real life problems (city traffic simulation, nationwide epidemics, neural networks, as well as business models) often require a mega number of agents to simulate. As such, agent-based models (ABMs) need to be able to populate these number of agents over a scalable simulation space in order to handle these problems. And although parallel and distributed simulation frameworks have slowly started to address these computational needs, non-computing scientists still tend to use GUI-rich, easy-to-use ABM interpretive platforms. This paper strives to identify the difficulties in the current parallel ABM frameworks and to propose any improvements that can be made. For these reasons, ABM applications were surveyed and modeled as seven benchmark tests. Agent and spatial descriptivity were analyzed as well their execution performance numbers. 

Monday, May 21

Kanishk Sharan

Chair: Dr. Marc Dupuis
Candidate: Master of Science in Cyber Security Engineering
1:15 P.M.; DISC 464
Automating Digital Forensics on Linux Environments

Most of the current forensic solutions for Linux systems are based on expensive commercial software tools. Existing research does not provide solutions that perform comprehensive hands-on start-to-end forensic solutions holistically. The primary focus of this project is to automate different phases of computer forensic investigation, focussing primarily on Linux based systems. The proposed solution integrates functionalities of open source ( and free of cost) command-line tools using bash and python scripts. Integral forensic processes such as disk imaging acquisition, volatile memory dump and information analysis are automated using customised scripts. This solution proposes timeline creation and incident reconstruction using MySQL database, as extracted data from the infected computer is inserted into separate database tables. The proposed forensic architecture discusses core aspects of Linux forensics in details and provides an end-to-end forensic solution for home computers against medium-level threats.

Tuesday, May 22

Anusha Prabakaran

Chair: Dr. Min Chen
Candidate: Master of Science in Computer Science & Software Engineering
1:15 P.M.; DISC 464
Online Review Credibility Analysis

Product reviews on online shopping sites have become the vital source of customers’ opinions. These reviews have a significant impact on purchasing decisions and product rankings on popular e-commerce websites. Unfortunately, for the desire of profit or fame, fraudsters (spammers) write deceptive reviews (spam reviews) appreciating or deprecating a product. These reviews mislead potential customers and negatively affect the revenue of many genuine organizations.  This fact has raised the need for an effective method to detect the fake reviews and spammers. Drawing from the literature, there are many types of spam detection methods that help to provide reliable resources to customers and businesses. Yet, these methods have drawbacks, like, the supervised approaches have imbalanced data, rating based filtering systems and linguistic approaches are impaired by shrewd spammers, and synthetic datasets do not match the real world scenarios. However, existing research does not use a combination of methods to detect the spam reviews. The aim of this project is to develop a practical end-to-end system that uses a set of three methods: detection of duplicate reviews, detection of anomaly in review count and rating distribution, and detection of incentivized reviews to analyze the Amazon review data and generate a score. This score indicates the credibility of the reviews of a product. The proposed system could facilitate businesses to identify and constrain on vendors and spammers engaging in these dishonest practices. This system could also aid in data mining and online spam filtering systems to filter the product reviews and refine the product rankings. These three methodologies complement each other and identified the spam products with greater accuracy by using a statistical credibility scoring system, without requiring significant computational resources, rather than using a single method.

Wednesday, May 23

Jeremy Albert

Chair: Dr. Kelvin Sung
Candidate: Master of Science in Computer Science & Software Engineering
1:15 P.M.; DISC 464
User-Centric Classification of Virtual Reality Locomotion Methods

The recent introduction of consumer grade virtual reality systems caused a renewed interest in the technology, but virtual reality locomotion remains an unsolved problem. Virtual reality locomotion is traveling in an infinite-scale virtual world, while remaining in the confines of a room-scale real world. This thesis proposes a classification framework of virtual reality locomotion methods to create a common platform to compare dissimilar methods and identify what an ideal method would be.

Virtual reality locomotion remains unsolved because of three fundamental challenges, nausea, presence, and fatigue. Nausea happens when there is a conflict between a user’s vestibular and visual senses; Presence is the level a user feels they are truly in the virtual world; Fatigue happens when a user expends muscular energy for a sustained period. Our framework classifies locomotion methods by the level these locomotion challenges are satisfied. Literature review and anecdotal experience suggests that to satisfy one challenge necessitates neglecting another.

To test the classification framework’s validity, 30 users tested the 3 most popular virtual reality locomotion methods in a common testbed. The users were questioned about their experience as it relates to nausea, presence, and fatigue. The results show that locomotion methods do experience a trade-off in satisfying the locomotion challenges, but this trade-off is not linear. This means that within classification framework, an ideal virtual reality locomotion method is possible.

Thursday, May 24

Saranya Krishnan

Chair: Dr. Min Chen
Candidate: Master of Science in Computer Science & Software Engineering
11:00 A.M.; DISC 464
Identifying Tweets With Fake News

In today’s world, online social media plays a vital role during real world events, especially crisis events. There are both positive and negative effects of social media coverage of events, it can be used by authorities for effective disaster management or by malicious entities to spread rumors and fake news. Recently many people expressed their concerns about the US elections-2016. There were many arguments and debates stating that unreliable news was published about the candidates that changed the election results. These types of information alters the results and creates false opinion among the general public. Taking into consideration the harmful consequences of false media content, there is a profound need to detect and control false information and prevent its spread. To address this issue, this capstone project presents a web system to identify tweets that tend to spread fake news and images using the following techniques:

1. Statistics of twitter user account

2. Machine learning model trained using J48 and SVM

3. Reverse Image Search

4. Cross verification with Megan Risdal dataset of fake news sources

5. A “crowd sourcing” approach

The results show that J48 performs better than SVM in predicting fake tweets. In our study, when we experimented training the classifier on one dataset and testing it on a completely different dataset, we found that accuracy was on the lower side. However the recall in predicting fake tweets was much higher which is the ultimate goal of this project.

Friday, May 25

Yun-tai Chang

Chair: Dr. Marc Dupuis
Candidate: Master of Science in Cyber Security Engineering
8:45 A.M.; DISC 464
A Two-layer Authentication Using Voiceprint for Voice Assistants

Voice assistants are a universal service in nowadays environment, they can be around with us for 24 hours through different devices such as mobile or hardware sitting in the house. More and more people are using voice assistants since voice assistants have more functions and abilities to help users in their daily life. However, the security of voice assistants does not increase as much as the rising population and mounting abilities. The lack of authentication mechanism in voice assistants gives attackers an opportunity to exploit voice assistants to control and get personal information from linked services. To protect voice assistants from attacks and keep the usability of voice assistants, this project proposes a voice authentication method. We utilize Microsoft cognitive speaker recognition API and Google speech API to implement an Android application to examine the approach. The result indicates that the voice authentication method can resist all kind of voice attacks and it is easy to use and learn for users.

Kevin Wu

Chair: Dr. Brent Lagesse
Candidate: Master of Science in Cyber Security Engineering
1:15 P.M.; DISC 464
Detecting Streaming Wireless Cameras with Timing Analysis

The concept of Internet of Things (IoT) is growing rapidly to provide convenience, as sensors collect, communicate, and collaborate with each other to provide better services. Wi-Fi cameras have been massively manufactured and widely adopted to provide monitoring service at fairly low cost. Although Wi-Fi cameras provide real-time monitoring service, those devices often come with weak security mechanisms. In the case of preventing abuse use of Wi-Fi camera spying and monitoring, we proposed a novel method to detect hidden Wi-Fi cameras with timing analysis. Further, a mobile phone is used as a detector to identify hidden Wi-Fi cameras. In order to provide constant and faster communication, IoT devices often required low-latency networks. The proposed methodology performs statistical analysis (Correlation Coefficient, Dynamic Time Warping, Kullback-Leibler divergence and Jensen-Shannon divergence) to measure similarity score between network traffic streams and the recorded video from the mobile phone. The similarity score is further adapted to identify hidden Wi-Fi cameras in the environment. The result showed that the proposed detection methodology can successfully discover hidden Wi-Fi cameras with an accuracy rate of 97.436%.

Wednesday, May 30

Emily Feiping Li

Chair: Dr. Wooyoung Kim
Candidate: Master of Science in Computer Science & Software Engineering
11:00 A.M.; DISC 464
A data analysis and visualization processing tool for Quantitative Multiplex Co-Immunoprecipitation(QMI) Platform

The project aims to provide an intuitive data analysis and interactive visualization processing program that facilitates the Quantitative Multiplex co-immunoprecipitation (QMI) platform. The QMI platform is a novel approach that generates protein-protein interactions to provide medical practitioners with necessary data for engineering T-Cells, that can help patients’ immune systems fight against abnormal cells. However, the existing QMI platform has poor usability that lacks visualization and interactive interfaces, which necessitates an integration of scattered tools into a unified graphical user interface program. Here, we build a graphical user interface (GUI) based QMI analysis software program, which provides a streamlined workflow and an intuitive and interactive interface by integrating three separate analytic programs into a signal program. We believe that the final produce will have a huge impact on medical researches that require complex data analysis processes.

Thursday, May 31

Athira Sunil

Chair: Dr. Min Chen
Candidate: Master of Science in Computer Science & Software Engineering
1:15 P.M.; DISC 464
Text-to-Image Synthesis: Creating Visual Representation of Natural Language Sentences

Generating the visual representation of textual information is a challenging yet interesting topic with many potential applications. For instance, the technology can be used for automatic generation of text illustration, language translation using images as intermediaries and more descriptive and intuitive sentence-based image search. Recently, there have been significant efforts to study the relationship between natural language sentences and their image representations. This study can be done in either direction. That is, an image can be given as input and a sentence can be produced as output or an animation or scene can be generated from a textual description.

In this paper, we propose a novel approach to visualize natural language sentences using ImageNet to enhance language education. Currently, the focus is to assist English language learners in building their vocabulary of common nouns and developing an in-depth understanding of the various prepositions of locations. To achieve this goal, real-world images representing nouns are obtained from ImageNet and their foreground objects of interest are extracted using image segmentation. The objects are then rearranged on a canvas based on their spatial relationship specified in the sentence. To demonstrate the effectiveness of the proposed approach, a prototype application is developed to create visual representations of natural language sentences to assist in learning new vocabulary and structures during language education.

 Jewel (Yun-Hsuan) Lee

Chair: Dr. Michael Stiber
Candidate: Master of Science in Computer Science & Software Engineering
3:30 P.M.; DISC 464
Machine Learning of Spatiotemporal Bursting Behavior in Developing Neural Networks

Experimental investigation of the collective dynamics in large networks of neurons is a fundamental step towards understanding the mechanisms behind signal and information processing in the brain. In the last decade, the emergence of high performance computing technology has allowed long-duration numerical simulations to model large-scale neural networks. These simulated networks exhibit behaviors (ranging from stochastic spiking to synchronized bursting) that are observed in living preparations. These simulations’ high spatiotemporal resolution and long duration produce data that, in terms of both quantity and complexity, challenge our interpretative abilities. This thesis presents an application of machine learning techniques to bridge the gap between microscopic and macroscopic behaviors and identify the small-scale activity that leads to large-scale behavior, reducing data complexity to a level that can be amenable to further analysis. 

Friday, June 1

Rashi Goyal

Chair: Dr. Erika Parsons
Candidate: Master of Science in Computer Science & Software Engineering
1:15 P.M.; DISC 464
Building a Machine Learning Based Recommendation Engine for the Virtual Academic Advisor System

Machine Learning is presently being used to tackle various problems from voice recognition to self-driving vehicles.  However, there are many areas where modern software applications have not reached due to lack of interest or budget.  The education sector is one of these areas, and itself poses a set of interesting questions suitable for applied Machine Learning and modern Data Analysis approaches, which can greatly benefit the community. A relatable example is the problem of a student choosing a career path, the basis of which is an appropriate academic plan. In our state, community college students have difficulty choosing a career path as they do not have a well-defined academic path to transfer to a university and major of their choice. This is due to the fact that most of the advising is done with archaic tools (if any), and faculty also often pose as academic advisors when they are already overwhelmed by their daily responsibilities.  Moreover, each student has specific preferences like the choice of school, major, budget, time preference, etc., making the task of generating the study plans burdensome.  Study plan creation is a form of scheduling problem, and it is not trivial. There is little research on scheduling algorithms that address the problem of finding and recommending multiple paths going from multiple starting points to multiple goals (e.g., building prerequisite networks).  The goal of this research is to help community college students and advisors by implementing a Machine Learning recommendation system that automates the selection of most suitable academic plans, specifically, to transfer to four-year institutions, based on personal preferences.

William Schneble

Chair: Dr. Geethapriya Thamilarasu
Candidate: Master of Science in Computer Science & Software Engineering
3:30 P.M.; DISC 464
Federated Learning for Intrusion Detection Systems in Medical Cyber-Physical Systems

Medical Cyber-Physical Systems (MCPS) are networked systems of medical devices with seamless integration of physical and computation components. MCPS are increasingly used in healthcare environments to deliver high quality care by enabling continuous monitoring and treatment. However, security breaches can compromise privacy, integrity, and availability for medical devices while circumventing traditional approaches such as cryptography. This can lead to severe repercussions for both the patient and hospital in terms of injury and liability. We implement a massively distributed, machine-learning-based IDS for the MCPS space based on Federated Learning -- FLIDS. We evaluate our design with real patient data and against Denial of Service (DoS), data modification, and data injection attacks. Our approach transmits 3.8 times fewer bytes than collecting the data at a central location which saves bandwidth. We also achieve a detection accuracy of greater than 99.0% and a False Positive Rate (FPR) of 1.0%. Lastly, we show that FLIDS can cope with unevenly distributed data and is a scalable solution that leverages the computing resources of many mobile devices.

Back to top

Questions: Please email cssgrad@uw.edu