Abstract
This portfolio brings together two projects that examine how existing computational
systems can be evaluated and improved to better serve human needs. The technical project
investigates whether machine learning methods applied to eye-tracking data can strengthen
support for ADHD assessment. The STS project examines how the growing difference between
the capability of anonymity infrastructure, specifically Tor, and public trust in this service affects
this privacy space in general. In both cases, the central issue is not simply technical capability,
but practical human value. A system may be powerful in theory, yet still fall short if it does not
generalize well, if its limits are poorly understood, or if the people who rely on it misjudge what
it can actually do. Together, these projects argue that computational systems contribute most
meaningfully to human betterment when they are not only innovative, but also aligned with the
needs of the people they affect.
The technical report, Benchmarking Gaze Biomarkers: A Cross-Dataset Evaluation of
Kinematic, Physiological, and Statistical Features for ADHD Detection, addresses the challenge
of developing more objective methods for ADHD classification. ADHD assessment still depends
heavily on clinical observation, behavioral judgment, and questionnaires, which remain valuable
but can introduce subjectivity and inconsistency. Eye-tracking offers a promising complementary
approach because it captures measurable behavioral and physiological signals related to
attention, cognitive control, and arousal. This project evaluates how well current state-of-the-art
computational approaches can support that goal. Specifically, it analyzes three prior methods: a
gaze-centered end-to-end sequence model, a pupil-centered imputation-aware deep learning
model, and a multimodal biomarker pipeline. Two of these models were reproduced and tested
on a public ADHD pupil dataset containing approximately 9,974 trial examples from about 50
subjects, using leave-one-subject-out evaluation and subject-level classification metrics. In the
current reproduction setting, both baselines performed weakly when transferred to this
pupil-centered dataset, indicating that published models may not generalize reliably when
preprocessing methods, data conditions, and training environments differ from the originals.
These results are important because they highlight the limitations of current systems rather than
assuming that prior reported performance will automatically transfer across datasets. Based on
this analysis, the project proposes a multimodal architecture that jointly models raw gaze and
raw pupil trajectories in a single end-to-end framework. The report therefore contributes both a
benchmark of existing approaches and a clearer rationale for how those approaches might be
improved to better support objective ADHD biomarker development.
The STS research paper, The Anonymity Gap: How Changing Threat Models Reshape
Trust, Risk, and Behavior on Tor, examines how an existing privacy system can become more
difficult to trust as the technical environment around it changes. Tor remains an important
infrastructure for journalists, activists, monitored populations, and others who depend on
anonymity for safety and privacy. However, recent machine-learning-based traffic-analysis
research has complicated the conditions under which Tor’s protections hold. The paper argues
that the key problem is not simply that new attacks exist, but that user trust can persist even as
real protections become more conditional and harder to interpret. This mismatch is described as
the “anonymity gap”: the distance between what users believe Tor safely provides and what their
actual threat environment may allow. To support this argument, the paper compares technical
literature on website fingerprinting and flow correlation, realism-oriented studies that question
how well those attacks transfer to real-world conditions, and user-centered research showing that
Tor is often learned socially through institutional practices, community advice, and partial mental
models rather than formal threat analysis. The paper also includes a bounded technical
experiment using a sampled public website-fingerprinting dataset. These findings together show
that anonymity infrastructure cannot simply be assumed to remain protective on its own; rather,
it must be continuously reassessed and improved if it is to keep serving the people who depend
on it most. Concurrently, the literature shows that this anonymity gap is a very real issue and
addressing it is a necessary function for those who rely on these services.
Taken together, the two projects present a shared argument about computational systems
and human betterment. The technical project focuses on improving the usefulness of machine
learning for diagnosis by identifying where current models fail to transfer and how multimodal
design may offer stronger support. The STS project focuses on improving the usefulness of
privacy infrastructure by showing how changing attack conditions can weaken protection when
trust remains based on outdated assumptions. In both cases, the work emphasizes that existing
systems are most valuable not when they are treated as complete solutions, but when they are
both interpreted and used most effectively in relation to real human consequences.
