Organizational Unit:

Undergraduate Research Opportunities Program

Permanent Link

https://hdl.handle.net/1853/70802

Parent Organization

Organizational Unit

Office of Undergraduate Education

ArchiveSpace Name Record

https://finding-aids.library.gatech.edu/agents/corporate_entities/1180

Full item page

Publication Search Results

Now showing 1 - 10 of 62

MACH2: System for Root Cause Analysis of Kernel Vulnerabilities

(Georgia Institute of Technology, 2021-12) Desai, Sidhesh

Kernel code is ubiquitous in the modern technology landscape, and therefore, enforcing its security is of high importance. A common problem among modern kernel fuzzers is the discovery of vulnerabilities whose causes are difficult to pinpoint, meaning that they cannot easily be patched by developers. This leads to a large accumulation of bugs for kernel and kernel driver code. This issue can be remediated by being able to trace the root cause of a given exploit in the original source code. This study introduces MACH2, a system through which kernel vulnerabilities can have their root causes pinpointed such that they can be easily corrected by developers and/or automated systems. The MACH2 system consists of a 2-stage process: first, the system generates a trace of the exploit being run, and then, it uses this trace in tandem with a DSE engine to find the input regions of the code corresponding to the vulnerability at hand. MACH2 has already demonstrated its usability against CVEs and real-world exploits, and with upcoming additions, will be able to handle a wide array of vulnerability classes, allowing for a more secure kernel code landscape.
Chainsaw Before Scalpel: Dependency-Based Pre-processing for Program Reduction

(Georgia Institute of Technology, 2021-12) Rousskov, Mark

Program reduction techniques, which aim to minimize the size of a program, have many applications, including software debloating, debugging, and optimization in general. Therefore, these techniques have been extensively studied for decades. Past work in this area has typically focused on either performing larger, more effective edits (e.g., Delta Debugging, Hierarchical Delta Debugging) or reducing the search space based on a language grammar (e.g., Perses, C-Reduce). Most of these techniques had a primary goal of minimal output size, with reduction speed only as a secondary goal. We propose Chainsaw, a novel approach that improves existing techniques by offloading a subset of the reduction to a pre-processing step. Since Chainsaw does not need to be as thorough as existing reducers, this creates an opportunity to take a new approach which can benefit overall end-to-end performance. Our key insight is that in practical application, a considerable amount of input code is not needed, and dependency analysis enables effective and fast identification of this removable code. This dependency analysis is both general, thus easily applicable to different languages, and inexpensive, thus amenable to a speedy pre-processing step. Such analysis can enable the higher-fidelity techniques previously developed to skip a significant quantity of work and produce better results more quickly. We also present a prototype tool based on our approach. Our tool finds unused sections of code by analyzing the dependencies between items in the input text and is straightforward to implement. We leverage existing analysis tooling via the Language Server Protocol to easily identify dependencies. Our initial results are promising and show that our approach is extremely fast and can yield up to twofold end-to-end speed improvement when used as a pre-processor with existing state-of-the-art techniques.
An Application for Urban Analytics

(Georgia Institute of Technology, 2021-12) Steinichen, Charlotte Jane

The objective of this research was to develop a new, data-based methodology for analyzing urban environments. By combining graph-based street network data with socioeconomic data scraped from open sources such as Google Places and Foursquare, the application designed for this study provides a quantitative understanding of the urban landscape surrounding stadium projects. The application has been shown to be flexible and can be applied to urban environments across the globe. As a result, this study is a promising first step towards a comprehensive, data-based urban model that can be used to assist place-making professionals both in understanding existing urban development and in siting new projects.
Optimal-horizon model-predictive control with differential dynamic programming

(Georgia Institute of Technology, 2021-12) Stachowicz, Kyle W.

We present an algorithm, based on the Differential Dynamic Programming framework, to handle trajectory optimization problems in which the horizon is determined online rather than fixed a priori. This algorithm exhibits exact one-step convergence for linear, quadratic, time-invariant problems and is fast enough for real-time nonlinear model-predictive control. We show derivations for the nonlinear algorithm in the discrete-time case, and apply this algorithm to a variety of nonlinear problems. Finally, we show the efficacy of the optimal-horizon model-predictive control scheme compared to a standard MPC controller, on an obstacle-avoidance problem with planar robots.
Investigating the Impact of Estimated Modalities in Multi-Modal Activity Recognition

(Georgia Institute of Technology, 2021-12) Rajan, Rahul

RGB-D data obtained from affordable depth-sensors, like the XBox Kinect has allowed for remarkable progress in the field of human activity recognition (HAR). Depth information has been found to significantly increase performance in HAR tasks, especially when it’s fused with other modalities like RGB and Optical flow. Unfortu- nately, the use of depth sensors limits where these models can be used since these sensors are often difficult to use in outdoor settings. Additionally, most videos available today are shot on traditional video cameras, which don’t provide depth information needed to run RGB-D based HAR models. Fortunately, deep learning has al- lowed us to estimate this depth data with high accuracy from just RGB video. This paper investigates the viability of directly using this estimated depth information in RGB-D models for HAR-related tasks.
Natural Language Processing with Assemblies of Neurons

(Georgia Institute of Technology, 2021-12) Jung, Seung Je Je

The Assembly Calculus is a novel framework intended to bridge the gap between the level of neuron and synapses, and that of cognition. The Assembly Calculus is a computational system entailing a basic data item called an assembly, a stable set of neurons explained below; a set of operations that create and manipulate assemblies; and an execution model which is squarely based on basic tenets of neuroscience. Importantly, it allows the creation of biologically plausible, flexible and interpretable programs, enabling one to develop tangible hypotheses on how specific brain functions may work. Here, to help lay groundwork for the creation of algorithms in this framework, we present a natural language processing algorithm to solve the analogy task. Further, to facilitate such experimentation, we present here a tool which in real-time allows the simulation, modification and visualisation of this computational system, including several prepared examples. Lastly, we also present empirical analysis of the capabilities of the assembly calculus to store information in brain areas.
Exploring Natural Language to Visualization translation through conversational interaction

(Georgia Institute of Technology, 2021-12) Mitra, Rishab

Natural language interfaces (NLIs) have allowed users to explore and visualize data without much overhead in learning how to operate the interface for visual analy- sis. However, the state-of-the-art NLIs for visualization only support one-shot queries with minimal capability for follow-up queries, so users are unable to build upon previ- ous visualizations. We introduce the Conversational Interaction for Data Visualization (CI4DV) toolkit for developers, a toolkit for natural language to data visualization that takes in a natural language query and dataset as input and outputs a visualization speci- fication. CI4DV offers the additional capability to follow-up on previous visualizations, allowing developers to build NLIs on top of CI4DV that allow users the opportunity for conversational interaction with the NLI. We also propose two applications that are built upon CI4DV to demonstrate the toolkit’s capabilities and usage - an application that features an editable mind map and a Jupyter notebook with nested cells.
Distributionally Robust Optimization Techniques for Stochastic Optimal Control

(Georgia Institute of Technology, 2021-12) So, Chun Man Oswin

Distributionally robust optimal control is a relatively new field of robust control that tries to address the issue of safety by hedging against the worst-cast distributions. However, because probability distributions are infinite-dimensional, this problem is in general computationally intractable. This thesis provides an overview of applications of distributionally robust optimization for stochastic optimal control. In particular, we look at existing and potentially new computationally tractable methods for performing distributionally robust optimal control using the Wasserstein metric.
Innovating Internet Connectivity in the Atlanta Westside Communities

(Georgia Institute of Technology, 2021-12) Smith, Aviva

There are thousands of people in the U.S. who do not have 24/7 internet access at their fingertips. To access broadband internet service, they must rely on the good will of neighbors, libraries, or other public venues that offer to access to affordable, stable internet connection. People facing this situation find themselves unable to submit homework assignments, apply for jobs, work from home, stay up to date with local and global news, or participate in social media. What has made modern life convenient for so many, remains an inconvenience for some. Furthermore, the limitations of the lock downs put in place to prevent the spread of COVID-19 in 2020 has further exposed that many people would be unable to adapt significant portions of their lives to be maintained using internet accessible tools. As many areas of life have begun to incorporate the usage of resources only accessible online, having access to broadband internet is no longer just a convenience, but a necessity. Previous research has analyzed how people around the world handle the challenges of having limited access to internet and discussed the measures they take to circumvent these difficulties. However, the effectiveness of proposed solutions varies on the circumstance analyzed in the study. The Westside neighborhood of Atlanta Georgia is a prime example of an urban community in which many of the residents face the challenge of finding stable internet connection in their daily lives. By analyzing U.S. census data illustrating the correlation between internet accessibility and income, we determined that the Westside could be categorized as a low-median income area in which many residents do not have broadband internet access in their home and have few public venues where they can find alternative means of connecting to internet. Through investigating how community members use the internet, determining what are their usual means of finding internet access, and learning how they circumvent the challenges, we have been able to prototype a practical solution that could be used to improve internet connectivity in the Westside.
Volatility Prediction using Earnings Call Transcripts and Rationale Behind the Model's Predictions

(Georgia Institute of Technology, 2021-12) Eidnani, Dheeraj Deepak

In this study, we looked at creating models that predict volatility given earnings call transcripts that perform better than baselines such as HTML. To accomplish this, we developed the Common Transformer Stack (CTS), which jointly predicted the most significant sentences in each earnings call and the resulting change in volatility, and we developed a dataset of annotated rationales for around 500 earnings calls. Experiments showed that CTS significantly outperformed over baselines, but this improvement was due to the model explicitly paying attention to a select number of sentences rather than the rationales that we annotated. Regardless, we believe that the model and training methodology has promise to further improve upon baselines if improved approaches to annotate the rationale of an earnings call are developed, which we hope to explore in the near future.