Author ORCID Identifier
0000-0003-0546-7577
Defense Date
2026
Document Type
Dissertation
Degree Name
Doctor of Philosophy
Department
Computer Science
First Advisor
Milos Manic
Second Advisor
Kostadin Damevski
Third Advisor
Preetam Ghosh
Fourth Advisor
Lulin Yuan
Fifth Advisor
Bjorn Vaagensmith
Abstract
Deep learning has emerged as a powerful and versatile modeling paradigm, achieving state-of-the-art performance across a wide range of domains and applications. However, its success is heavily dependent on the availability of large amounts of high-quality training data, which is often expensive, time-consuming, or impractical to obtain. As a result, training datasets frequently provide only limited coverage of the underlying domain.
When data is insufficient, deep learning models can exhibit two distinct failure modes. The first is a privacy failure, in which the model memorizes instance-specific information that can later be extracted by an adversary. The second is a generalization failure, in which the model learns patterns that do not hold beyond the training data. Although these failures appear different, both stem from the same underlying problem: underdetermination. When the available data is insufficient to uniquely constrain a solution, multiple hypotheses can explain the training observations equally well, allowing optimization to converge to solutions that rely on memorization or spurious correlations rather than the true structure of the domain.
This dissertation argues that incorporating prior knowledge offers an effective way of addressing these limitations. By constraining the hypothesis space beyond what limited data alone can support, prior knowledge guides learning toward solutions that better reflect the underlying structure of the problem. We organize prior knowledge based on the mechanism through which it acts: excluding undesirable solutions, transferring knowledge from related data-rich settings, or explicitly encoding domain structure into the model. Each mechanism is evaluated in a setting marked by a distinct form of data scarcity: within-subject sample scarcity, annotation scarcity, and limited behavioral diversity.
First, we address the privacy failure arising from within-subject sample scarcity in EEG-based brain-computer interfaces. We show that membership inference attacks can identify training participants even under black-box access, and that an exclusion prior enforcing participant-invariant representations reduces privacy leakage while maintaining task performance.
We then address the generalization failure in two domains with different forms of data scarcity. In radiotherapy dose planning, where expert annotations are limited, we develop a diffusion-based framework conditioned on a transferred prior from a pretrained medical segmentation model. By using embeddings extracted from the pretrained segmentation model in place of manually annotated organs-at-risk (OAR) masks and incorporating global context from a 3D Computed Tomography (CT) encoder, the framework achieves competitive dose prediction without requiring segmentation at inference time. In Controller Area Network (CAN) bus anomaly detection, where datasets often exhibit limited behavioral diversity, we incorporate an engineered prior derived from protocol structure into both the model architecture and the anomaly-scoring process, resulting in substantial improvements in detection performance across multiple datasets.
Together, these results demonstrate that aligning prior knowledge with the particular form of data scarcity improves both the privacy and the generalization of deep learning models in data-constrained applications.
Rights
© The Author
Is Part Of
VCU University Archives
Is Part Of
VCU Theses and Dissertations
Date of Submission
8-3-2026