Computer Science Theses and Dissertations
Permanent URI for this collection
Now showing 1 - 5 of 910
- ItemReal-Time Cybersecurity Situation Awareness Through a User-Centered Network Security Visualization(2022) DeValk, Kaitlyn; Elmqvist, Niklas; Computer Science; Digital Repository at the University of Maryland; University of Maryland (College Park, Md.)One of the most common problems amongst cybersecurity defenders is lack of network visibility, leading to decreased situation awareness and overlooked indicators of compromise. This presents an opportunity for the use of information visualization in the field of cybersecurity. Prior research has looked at applying visual analytics to computer network defense, which has led to the development of visualizations for a variety of use cases in the security field. However, many of these visualizations do not consider user needs and requirements or require some predetermined user knowledge about the network to create the visuals, leading to low adoption in practice. With this in mind, I took a bottom-up, user-centered approach using interviews to gather user-desired components for the design, development, and evaluation of a network security visualization tool, called Riverside. I designed a visualization that attempts to balance providing a comprehensive view of an environment while supplying details-on-demand. Riverside’s key contribution is a data-driven, dynamic view of a network’s security state over time, meant to supplement an analyst’s real-time situation awareness of their network. Riverside’s system automatically partitions internal from external network components to visualize potential attack vectors across the entire environment. This research supports the need for further incorporation of users into the cybersecurity visualization development lifecycle. I call attention to key requirements for creating effective cybersecurity visualizations and specific use cases where visualizations can be leveraged to augment operational cybersecurity capabilities.
- ItemROBUSTNESS AND UNDERSTANDABILITY OF DEEP MODELS(2022) Ghiasi, Mohammad Amin; Goldstein, Thomas; Computer Science; Digital Repository at the University of Maryland; University of Maryland (College Park, Md.)Deep learning has made a considerable leap in the past few decades, from promising models for solving various problems to becoming state-of-the-art. However, unlike classical machine learning models, it is sometimes difficult to explain why and how deep learning models make decisions. It is also interesting that their performance can drop with small amounts of noise. In short, deep learning models are well-performing, easily corrupted, hard-to-understand models that beat human beings in many tasks. Consequently, improving these deep models requires a deep understanding. While deep learning models usually generalize well on unseen data, adding negligible amounts of noise to their input can flip their decision. This interesting phenomenon is known as "adversarial attacks." In this thesis, we study several defense methods against such adversarial attacks. More specifically, we focus on defense methods that, unlike traditional methods, use less computation or fewer training examples. We also show that despite the improvements in adversarial defenses, even provable certified defenses can be broken. Moreover, we revisit regularization to improve adversarial robustness. Over the past years, many techniques have been developed for understanding and explaining how deep neural networks make a decision. This thesis introduces a new method for studying the building blocks of neural networks' decisions. First, we introduce the Plug-In Inversion, a new method for inverting and visualizing deep neural network architectures, including Vision Transformers. Then we study the features a ViT learns to make a decision. We compare these features when the network trains on labeled data versus when it uses a language model's supervision for training, such as in CLIP. Last, we introduce feature sonification, which borrows feature visualization techniques to study models trained for speech recognition (non-vision) tasks.
- ItemA Multi-Faceted Approach for Evaluating Visualization Recommendation Algorithms(2022) Zeng, Zehua; Battle, Leilani; Computer Science; Digital Repository at the University of Maryland; University of Maryland (College Park, Md.)Data visualizations allow analysts to quickly understand data trends, outliers, and patterns. However, designing the "best" visualizations for a given dataset is complicated. Multiple factors need to be considered, such as the data size, data types, target analysis tasks being supported, and even how the visualization needs to be personalized to the audience. In response, many visualization recommendation algorithms are being proposed to reduce user effort by automatically making some or all of these design decisions for analysts. However, existing visualization recommendation algorithms are evaluated in isolation, or the comparisons do not measure user performance. In other words, existing algorithms are not tested in a way that aligns with how they are used in practice. The lack of evaluation approaches makes it impossible to know how functional an algorithm is compared to another across various analysis tasks, hindering our ability to design new algorithms that provide significantly more benefits than the existing ones.This dissertation contributes a multi-faceted approach for evaluating visualization recommendation algorithms to investigate factors affecting an algorithm's performance and ways to improve it. It first proposes an evaluation-focused framework and then demonstrates how the framework can evaluate strategic behaviors and user performance among a broad range of existing algorithms. The case study results show that newly proposed algorithms might not significantly improve user performance. One way to improve the algorithm performance is by integrating more established theoretical rules or empirical results on how people perceive different visualization designs, i.e., graphical perception guidelines, to guide the recommendation ranking process. Thus, this dissertation next presents a thorough literature review of existing graphical perception literature that can inform visualization recommendation algorithms. It contributes a systematic dataset that collates existing theoretical and experimental visualization comparison results and summarizes key study outcomes. Further, this dissertation conducts an exploratory analysis to investigate the influence of each piece of graphical perception study in changing a visualization recommendation algorithm's behavior and outputs. The analysis results show that some graphical perception studies can alter the behavior of visualization recommendation algorithms dominantly, while others have little influence. Based on the analysis findings, this dissertation opens avenues at the intersection of graphical perception and visualization research, like how to evaluate the effectiveness of new graphical perception work in guiding visualization recommendations.
- ItemSENSITIVITY ANALYSIS AND STOCHASTIC OPTIMIZATIONS IN STOCHASTIC ACTIVITY NETWORKS(2022) Wan, Peng; Fu, Michael C; Applied Mathematics and Scientific Computation; Digital Repository at the University of Maryland; University of Maryland (College Park, Md.)Activity networks are a powerful tool for modeling and analysis in project management, and in many other applications, such as circuit design and parallel computing. An activity network can be represented by a directed acyclic graph with one source node and one sink node. The directed arcs between nodes in an activity network represent the precedence relationships between different activities in the project. In a stochastic activity network (SAN), the arc lengths are random variables. This dissertation studies stochastic gradient estimators for SANs using Monte Carlo simulation, and the application of stochastic gradient estimators to network optimization problems. A new algorithm called Threshold Arc Criticality (TAC) for estimating the arc criticalities of stochastic activity networks is proposed. TAC is based on the following result: given the length of all arcs in a SAN except for the one arc of interest, that arc is on the critical path (longest path) if and only if its length is greater than a threshold. By applying Infinitesimal Perturbation Analysis (IPA) to TAC, an unbiased estimator of the derivative of the arc criticalities with respect to parameters of arc length distributions can be derived. The stochastic derivative estimator can be used for sensitivity analysis of arc criticalities via simulation. Using TAC, a new IPA gradient estimator of the first and second moments of project completion time (PCT) is proposed. Combining the new PCT stochastic gradient estimator with a Taylor series approximation, a functional estimation procedure for estimating the change in PCT moments caused by a large perturbation in an activity duration's distribution parameter is proposed and applied to optimization problems involving time-cost tradeoffs. In activity networks, crashing an activity means reducing the activity's duration (deterministic or stochastic) by a given percentage with an associated cost. A crashing plan of a project aims to shorten the PCT by reducing the duration of a set of activities under a limited budget. A disruption is an event that occurs at an uncertain time. Examples of disruptions are natural disasters, electrical outages, labor strikes, etc. For an activity network, a disruption may cause delays in unfinished activities. Previous work formulates finding the optimal crashing plan of an activity network under a single disruption as a two-stage stochastic mixed integer programming problem and applies a sample average approximation technique for finding the optimal solution. In this thesis, a new stochastic gradient estimator is derived and a gradient-based simulation optimization algorithm is applied to the problem of optimizing crashing under disruption.
- ItemEgocentric Vision in Assistive Technologies For and By the Blind(2022) Lee, Kyungjun; Kacorri, Hernisa; Computer Science; Digital Repository at the University of Maryland; University of Maryland (College Park, Md.)Visual information in our surroundings, such as everyday objects and passersby, is often inaccessible to people who are blind. Cameras that leverage egocentric vision, in an attempt to approximate the visual field of the camera wearer, hold great promise for making the visual world more accessible for this population. Typically, such applications rely on pre-trained computer vision models and thus are limited. Moreover, as with any AI system that augments sensory abilities, conversations around ethical implications and privacy concerns lie at the core of their design and regulation. However, early efforts tend to decouple perspectives, considering only either those of the blind users or potential bystanders. In this dissertation, we revisit egocentric vision for the blind. Through a holistic approach, we examine the following dimensions: type of application (objects and passersby), camera form factor (handheld and wearable), user’s role (a passive consumer and an active director of technology), and privacy concerns (from both end-users and bystanders). Specifically, we propose to design egocentric vision models that capture blind users’ intent and are fine-tuned by the user in the context of object recognition. We seek to explore societal issues that AI-powered cameras may lead to, considering perspectives from both blind users and nearby people whose faces or objects might be captured by the cameras. Last, we investigate interactions and perceptions across different camera form factors to reveal design implications for future work.