Prayaag Venkat and Ashton Webster awarded Honorable Mention by CRA for Research

Descriptive Image for Prayaag Venkat and Ashton Webster awarded Honorable Mention by CRA for Research

Continuing a tradition of excellence in undergraduate research in Computer Science, Prayaag Venkat and Ashton Webster were awarded Honorable Mention by the Computing Research Association for their undergraduate research projects on December 9th 2016.  Venkat conducted research with Professor Samir Khuller and Professor Dave Mount.  Webster conducted research with Professor Jim Purtilo and Professor Michel Cukier (of Mechanical Engineering and Electrical and Computer Engineering).   

While they were preparing for exams, Venkat and Webster kindly corresponded about their research that they conducted over the last year.  Venkat plans to apply to graduate schools in the next year in Theory, and Webster is currently in the BS/MS program here in the Department of Computer Science. 

Prayaag Venkat (Junior)

During the summer and academic year of 2014, I worked with Professor David Mount. Professor Mount and I designed space efficient data structures for solving fundamental problems in computational geometry. Often times, in application areas such as machine learning and physics, there is a large set of points and one wishes to query its geometric properties. These problems are well studied and usually solved by creating a data structure, such as a quadtree, that represents the point set in a way that allows queries to be answered efficiently and accurately. However, it is often the case that the original data set is so large that there is not enough space to create and store such a data structure. For example, in practical applications, one may wish to store the entire data and data structure in main memory, which will give faster processing but is limited in capacity. Professor Mount and I designed a new variant of the quadtree, and showed that it uses (close to) the information theoretic minimum number of bits required to answer certain geometric queries approximately.  This work was published and presented at CCCG 2014 (A Succinct, Dynamic Data Structure for Proximity Queries on Point SetsPrayaag Venkat and David M. Mount, pp. 216-225).

During the summer of 2016, I began a research project with Professor Samir Khuller as part of the University of Maryland CAAR REU (Research undergraduate experience). My group members and I developed approximation algorithms for scheduling problems. In modern datacenters, technologies like MapReduce divide large problems into small jobs, which need to be scheduled on complex architectures. My group introduced a general framework that converts algorithms for offline scheduling and knapsack-type problems (in which the entire input is known upfront) to the online setting (i.e., the entire problem input is not available upfront). We used our framework to produce algorithms that have the best known approximation performance for a variety of scheduling problems. Finally, we implemented our algorithms and found that they perform well on real world data sets. This work is currently in submission.  

 

Ashton Webster (BS/MS program):

I began working Professor Michel Cukier and a team of undergraduate and graduate students as a sophomore on a method for developing network scan detection models.  The paper we wrote describing our work was accepted to the SECURECOMM conference and we presented our work in Dallas, Texas in October 2015.  This work specifically Designs and improves a model to detect network port scans.  Port scans often appear before widespread attacks as attackers gather data, but they are rarely detected or investigated due to the challenges of creating a model capable of reliably identifying them.  I lead a team of two other undergraduate students and one graduate student as we created a general and practical framework that could create a “Local Optimal Model”, which is the best possible model for a given performance metric.  By processing millions of network flow records into a feature space for machine learning, we were able to create and analyze the performance of thousands of different detection models.  I personally worked on the implementation of the experiment and analyzed most of the results by creating various scripts and programs.

Since then, I have been working with Professor James Purtilo on using machine learning technique known as "transfer learning” to better identify software defects and vulnerabilities in source code.  This semester I completed my honors thesis on this topic. I have studied another applied machine learning task: identifying vulnerabilities and defects in software projects by analyzing source code metrics and tokens. I designed experiments to test if it is possible, under certain conditions, to use transfer learning techniques for this task; specifically, to determine if it is possible to use one or more vulnerability- or defect-labeled project to predict another, separate project. This work was particularly challenging due to lack of available implementations of transfer learning methods (almost no publicly available implementations exist) and computationally intensive model construction and evaluation (many experiments lasted many days and sometimes weeks). The results are still in the preliminary stage, but they indicate that, in general, no method performs better than the most naïve baseline: simply using all of the available source project(s) data to identify issues in the target project, without any sort of filtering or weighting. This result has significant implications suggesting that developers can identify a large proportion of the defects and vulnerabilities in the code without having to label any of their project’s files as vulnerable or defective, or even perform any complex filtering; they can simply use existing, potentially dissimilar, labeled projects. I also contributed open-source implementations for several different transfer learning algorithms as defined in other research papers, which will make it significantly easier for practical and academic use of these models.

List of CRA winners

2017   Prayaag Venkat, Honorable Mention (advisors: Samir Khuller and Dave Mount)

           Ashton Webster, Honorable Mention (advisors: Jim Purtilo and Michel Cukier)

2016:  Andrea Bajcsy, Honorable Mention (advisor: Yannis Aloimonos)

            Frank Cangialosi, Honorable Mention (advisor: Dave Levin)

2013:  Victoria Lai, Finalist (advisor: Dave Levin)

2012:   Elissa Redmiles, Honorable Mention (advisor: Samir Khuller)

2010:  Matt McCutchen, Awardee

           Allison (Hoch) Janoch, Honorable Metion

           John Silberholz, Honorable Mention

2008:  Katrina LaCurts, Honorable Mention

2007:  Jessica Chang, Honorable Mention

 

The Department welcomes comments, suggestions and corrections.  Send email to editor [-at-] cs [dot] umd [dot] edu.