Our Ph.D. students work hard for years honing their research. This effort culminates in an oral dissertation where they present, explain, and defend their ideas. The final defense of the dissertation is conducted by an examination committee recommended by the graduate faculty advisor and approved by the dean of The Graduate School. All final oral examinations are open to members of the graduate faculty.
Below is the published list of upcoming or recent dissertation defenses for Information Systems students.
Announcement of Ph.D. Dissertation Defense
First, we investigated experts’ (n=20) understanding of mobile security and how their concerns shaped everyday use of those platforms. They experienced typical usability problems and situational impairments with their mobile devices, but also described caution towards data sharing. The avoidance was based on factors including the sensitivity of the data to be shared and variable distrust of underlying network connections and technology platforms. We then secondly compared the outlook and behavior of a similar cohort (n=38) of both experts and non-experts when considering adoption of biometric authentication on their mobile devices. Experts were found to more readily accept fingerprint unlocking as an improved mode of authentication than non-experts. However, experts resisted its use for sensitive transactions, and their enthusiasm did not transfer to facial recognition. Our third study examined perceptions of experts, in two rounds (n=8, and n=19), serving security consultant roles (CISOs) for small businesses. CISOs tended to view government-authored security guidance as harder to use but more authoritative than commercial sources, and saw small businesses as highly vulnerable to online threats.
This work concludes by comparing the observations and implications drawn from these studies to examples of security guidance. Guidance of this type was described as a key reference and basis for advisement by CISOs to non-expert small business owners. Based on the deductive comparison, a preliminary set of guidelines for presenting models of online risk are offered. These guidelines indicate ways to make security guidance more effective, based on experts’ perspectives that should be broadly applicable and useful as awareness of online threats becomes more prevalent.
Student: Sreenivasan Ramasamy Ramamurthy
Date/Time: Friday, May 14, 2021, 10:00 AM – 12:00 PM
Meeting number:120 459 9123
Dr. Bivas Mitra, IIT Kharagpur
The design, development, and deployment of human activity recognition (HAR) and personalized health monitoring applications in smart home environments have been accelerated with the recent penetration of wearable and off-the-shelf consumer-grade Internet-of-Things (IoT) devices, along with the advancement of artificial intelligence (AI) and machine learning (ML) techniques. In recent years, substantial research effort has been conducted to improve human activity recognition and physiological health monitoring using various deep learning paradigms (e.g. supervised, unsupervised, semi/self-supervised learning algorithms) and bio-signals (e.g. photoplethysmograph, electrocardiogram, galvanic skin response, electrodermal activity, etc). However, early detection of functional and behavioral health deteriorations due to the cognitive health decline using only the accelerometry data has been lightly investigated. In this thesis, we propose to study the relationships between activities and cognitive health declines for older adults longitudinally using accelerometry signals while focusing on addressing the challenges arising from micro- and macro- activity sensing, user-agnostic activity model learning, and personalized health monitoring.
To support this effort, we performed an extensive data collection drive to gather sensor-based activity data and survey-based clinical assessment on functional, behavioral, and cognitive health from 25 older adults residing in their own homes in a retirement community center. This collection of real data using our in-house smart home sensing systems from community-dwelling older adults helped us to realize the challenges associated with the missed detection of specific activities during the sensing phase, appropriate demarcation of activity transition points, availability of abundant unlabeled data, and severe paucity of labeled data. To address the above challenges, particularly the availability of abundance unlabeled data in a practical setting, first, we propose a self-taught learning approach that helps to increase the precision and reduce the missed detections in activity recognition and sensing, respectively. To further improve the activity recognition performance while addressing the demarcation of the activity transition points across multiple older adults, we propose a user-invariant scalable activity recognition framework relying on contrastive learning. Next, we propose a multi-task learning framework that helps exploit the correlations between the fine-grained functional and cognitive health decline measurement metrics.
Our proposed self-supervised activity and cognitive health monitoring framework relying on self-taught, contrastive, and multi-task learning approaches helps track the health decline accurately; nonetheless, we hypothesize that it needs to dynamically incorporate the evolution of psychosocial factors as the older adults age while parameterizing the neurodegenerative disease progression as an indicator of overall health decline. The functional health changes (e.g. wandering, incorrect sequence, etc,) manifested through behavioral symptoms (agitation, depression) during the early stages of dementia are subtle but still can be quantified through the interruption and deviation from specific activities. The stages of dementia with respect to cognitive impairments e.g., no, mild, moderate, and severe cognitive decline are labeled across the distributions of data conditioned on the user using survey and observation-based assessments in a clinical setting as opposed to the user’s current and past activities. The latter is practically expensive due to the challenges in longitudinal sensing and data collection including the appropriate labeling of the data. To address this challenge, we propose a reinforcement learning-based framework that leverages the placement of the representations learned via contrastive loss and the clinician input to model a reward function that enables us to integrate the psychosocial factors with the neurodegenerative disorder indicator. We evaluated our framework on 25 older adults dataset collected from a retirement community center with IRB approval in place. Our proposed self-taught learning approach reduces the activity missed detection by 20%, the contrastive learning approach obtains an F1-score of 92% in detecting demarcation of overlapped activities, and the multi-task learning approach achieves an F1-score of 95% in classifying cognitive impairments concurrently with activities. Next, we plan to evaluate our reinforcement learning approach using our own and public datasets. Finally, we plan to build a mobile application, that connects with smart home IoT devices, harnesses data, executes the proposed algorithms for health assessments, and provides just-in-time feedback about the current cognitive health status to both the user and clinicians along with the explanation for the recommended assessment.
Announcement of Ph.D. Proposal Defense
Student: Hirak Ray
Date/Time: Friday May 14th, 2021, 10am-1pm Eastern Time
Dr. Ravi Kuber, UMBC (Co-Chair)
Dr. Adam J. Aviv, George Washington University (Co-Chair), (firstname.lastname@example.org)
Dr. Foad Hamidi, UMBC
Dr. Yaxing Yao, UMBC
Dr. Florian Schaub, University of Michigan (email@example.com)
While researchers have examined privacy among a wide range of users, there has been lesser focus on the perceptions of older adults, who may exhibit attitudes towards privacy that differs from their younger counterparts. Lower levels of awareness regarding potential online privacy violations, coupled with limited knowledge of protective measures that can be adopted to counter online attacks lead to negative outcomes for older adults, including falling victim to scams and data breaches.
My research investigates older adults’ privacy and security perceptions regarding digital and non-digital technologies, and identifies how their usage and adoption of tools and technologies are impacted by these perceptions. To this end, two studies have been conducted, exploring (1) complex privacy behaviors of older adults and comparisons with younger age groups, and (2) adoption barriers and motivators of online security tools.
Currently, older adults are experiencing a higher amount of social isolation than before due to the social distancing requirements of the COVID-19 lockdown. For my dissertation, I intend to investigate the forced adoption of online tools among older adults due to this phenomenon, specifically towards online privacy and security issues which may surface, and see how this compares to adoption under less constrained circumstances. Findings from the work would complement findings from my prior studies to aid design guidance provided to interface developers and researchers aiming to support privacy among older groups of users.
Title: Understanding and Supporting Collaboration Between Youth in Technical Workplaces
Student: William Easley
Time: Friday, September 11, 2020, 12 pm – 3 pm
Location: Meeting Link: https://umbc.webex.com/umbc/j.php?MTID=mc3c17c758cbb6d0a640de365bf203ccc
Meeting number (access code): 120 111 8670
Meeting password: 3exRkqNYu34
Dr. Foad Hamidi (Chair/Advisor)
Dr. Helena Mentis (Chair/Advisor)
Dr. Anita Komlodi
Dr. Amy Hurst
Dr. Wayne G. Lutters
Dr. Betsy DiSalvo
The three studies included in this document all seek to better understand how youth coordinate work in technical settings with a specific focus on their usage of communication technologies. The first documents and unpacks instances where youth successfully and unsuccessfully coordinated their work during asynchronous handoffs. The second study characterizes the youths’ use of Slack, a popular workplace chatting tool and investigates the factors motivating and limiting adoption of this tool. The final study examines the youths’ attitudes and perceptions towards workplace communication technologies. Findings and recommendations from this dissertation offer insight into how we can better prepare and support our future workforce.
Title: Interpretable deep learning models for Electronic Health Records
Student: Peichang Shi
Advisor: Dr. Aryya Gangopadhyay
Time: Monday, June 22, 2019, 10 am – 12 pm
Deep learning neural networks (DNNs) have been proved to be powerful tools in computer vision and machine learning across robotics, healthcare, and security. However, despite their superior performance to traditional machine learning methods, it remains challenging to understand their inner mechanism and explain their downstream output due to the black box effects.
In this dissertation proposal, we aim to unfold the black-box and provide explanations in a transparent way using Electronic Health Records (EHRs). Our proposed approach is to add a partially connected layer after the output from the DNNs based on a decision tree. This decision tree layer could interpret DNNs through rules from the input vectors. Our method could achieve global interpretability for all the test instances on the hidden layers without sacrificing the accuracy obtained from the original deep learning model. This means our model is faithful to the original deep DNN model, which leads to reliable interpretations. Beyond that, we would also like to address some common issues related to EHRs including irregular time series data, censored data, multitask learning issues and hierarchical structure variables in the model.
The methods proposed in this dissertation will be validated with the End Stage Renal Disease (ESRD) Population Public Use File (PUF). This PUF data contains the current and historical ESRD patient population and associated clinical data, which has about 3.2 million patients including hospital admissions, hospitalizations, chronic conditions and 13 monthly clinical measures.