THESIS DEFENSE

Author :  Jakub Jerzmanowski
Advisors: Dr. Homayoun Valafar
Date: July 10, 2026
Time: 10:00 Am
Location:  Room 2265, Storey Innovation building

Link:  https://teams.microsoft.com/l/meetup-join/19%3ameeting_NWQzZDZhOTItMjNi…

Abstract
Cigarette smoking remains the leading cause of preventable death in the United States, claiming approximately 480,000 lives annually. Clinicians who tailor cessation interventions are limited by their data: knowing that a person smoked is far less useful than knowing when each puff began and ended, information from which count, duration, and inter-puff interval follow. Wrist-worn inertial measurement unit (IMU) systems can sense smoking unobtrusively, but existing methods classify coarse fixed-length windows rather than the puffs themselves, and so cannot recover this structure. We frame puff detection as dense, per-timestep segmentation, labeling every IMU sample and evaluating at the event level. To our knowledge, this is the first treatment of smoking-puff detection as a segmentation task. Using leave-one-subject-out cross validation, we show that window classifiers, run densely, collapse under strict overlap (event F1 at IoU 0.75 near 0.05), placing predictions in the right neighborhood but the wrong shape, whereas a 1D U-Net adapted to the time domain reaches 0.714 on our 1,500-hour, six-participant in-situ dataset. Decomposing the residual error shows that localization is essentially solved (matched-puff IoU of 0.93 to 0.99); the entire remainder is missed detections, concentrated in one hard participant who drives a pooled miss rate of 0.101. Because the bottleneck is per-person recall, we close it with personalization: warm restarting on as few as ten of a participant’s own puffs cuts the hard subject’s miss rate from 0.38 to 0.05 while boundary quality holds, reframing the problem as one of data and personalization rather than architecture.

DISSERTATION DEFENSE

Author :  Jonathan Sharp
Advisors: Dr. Csilla Farkas
Date: July 09, 2026
Time: 09:00 Am
Location:  Virtual
Link:  https://teams.microsoft.com/meet/22071693515848?p=LRAW7BhvdNDjot6moR

Abstract
In this dissertation, we studied how to improve multi-factor authentication using context-aware and adaptive authentication methods. We developed the ContextAware Adaptive Multi-factor Authentication (CAA-MFA) framework to enhance the usability and security of authentication systems in dynamic environments. Traditional multi-factor authentication (MFA) systems often rely on static combinations of factors regardless of contextual risk. This limits their effectiveness against evolving threats such as phishing, social engineering, credential compromise, and MFA
fatigue attacks (1). CAA-MFA addresses these challenges by adjusting authentication requirements based on real-time context, trust, policy constraints, and access risk.
The proposed framework treats authentication as a decision-making problem shaped by dynamic risk and policy compliance. It models user and environmental context semantically, evaluates the trustworthiness of available authentication factors, selects
authentication factors through constraint solving, and quantifies access risk using a Risk Level Assessment (RLA) model. By modeling context through ontologies and enforcing factor-selection constraints formally, CAA-MFA supports structured adaptation and scalable policy enforcement across heterogeneous environments (2). The framework also builds on trust-based reasoning for adaptive authentication by assigning trust scores to authentication factor-source pairs and selecting factors that satisfy constraints related to trustworthiness, privacy, usability, and required security level
(3).
The framework was evaluated using 12,000 labeled login attempts, including
iii
10,000 legitimate login attempts and 2,000 attacker attempts. The evaluation measured authentication performance, computational overhead, runtime cost, and the contribution of SAT-based factor selection and RLA. The strongest configuration,using both SAT-based selection and RLA with an SVM classifier, achieved an Equal Error Rate (EER) of 0.0102, Area Under the Curve (AUC) of 0.9965, and F1-score of 0.9852. These results show that CAA-MFA can distinguish legitimate and adversarial login attempts with strong performance while providing a structured method
for adjusting authentication strength according to contextual risk. The main research questions in this dissertation are as follows:

1. How can we model user and environmental context to support adaptive multi-factor authentication?
 

We proposed a semantic context model that captures relevant features from users, devices, behavior, history, and the surrounding environment. These features include user roles, device attributes, network conditions, location, time, behavioral indicators, and privacy requirements. The model uses ontologies to support structured reasoning about contextual conditions, enabling the authentication system to interpret contextual changes and supply meaningful inputs to downstream trust evaluation and factor selection.

2. How can we dynamically select authentication factors using constraintsolving based on contextual trust?
 

We proposed a formal mechanism for selecting authentication factors using trust evaluation and constraint satisfaction. The framework supports both passive and active authentication factors. Passive factors include device identifiers, location data, application history, and other contextual signals, while active factors include biometric input, typing behavior, user-entered PINs, and other explicit verification methods. Each factor-source pair is assigned a trust score iv reflecting reliability, source integrity, and contextual relevance. The selection of factors is modeled as a constraint satisfaction problem, and a SAT solver is used to enforce policy requirements such as minimum trust thresholds, usability constraints, and privacy constraints. This approach enables dynamic adaptation to changing contexts without relying on static authentication workflows.
 

3. How can we quantify the cost versus benefit of utilizing CAA-MFA over traditional MFA?
 

We evaluated the practical trade-offs introduced by deploying CAA-MFA compared to static MFA systems. The evaluation considers usability, computational overhead, implementation complexity, and security effectiveness. The results show that CAA-MFA introduces additional operational cost through context modeling, SAT-based factor selection, RLA computation, and classifier evaluation. However, these costs are justified in environments where authentication risk varies across users, devices, networks, and resources because the adaptive model provides measurable improvement in authentication performance and policy-aware factor selection.
 

4. How can we quantify access risk and use it to adjust authentication strength in real time?

We developed an access-risk model that incorporates contextual factors, authentication strength, historical user behavior, user clearance, and resource sensitivity. The model generates a continuous Risk Level Assessment (RLA) score that supports real-time adjustment of authentication strength. This score helps determine when to escalate verification requirements in high-risk contexts and when to reduce unnecessary authentication burden in low-risk contexts. The RLA model is integrated with the context-aware factor selection framework and evaluated through the authentication performance and ablation studies.

 DISSERTATION DEFENSE

Author :  Moh. Sabbir Saadat
Advisors: Dr. Sanjib Sur
Date: July 02, 2026
Time: 09:45 Am
Location:  Room 2277, Storey Innovation building

Link:  https://teams.microsoft.com/meet/29819819216139?p=qjSmXRE3Pm9C9PUJ2P

Abstract
Contactless perception systems are an attractive paradigm for applications in security or health monitoring due to their non-intrusiveness and ease of instrumentation compared to wearable or contact-based sensors. This dissertation advances the usability of contactless perception along two connected directions: millimeter-wave (mmWave)-based sensing and imaging, and multimodal digital health assessments.

MmWave signals are high-frequency (24.0 – 300.0 GHz) radio signals that can operate in low visibility, sense through some occlusions, avoid direct camera-like visual capture, and can sense micro-motions that are not possible with conventional vision systems (monocular camera, stereocamera, lidar, IR sensors). Moreover, mmWave signals constitute an important enabling layer in 5Gand-beyond networking paradigms, particularly in indoor networking scenarios. This presents us with an opportunity to bring mmWave sensing for applications where conventional vision systems fail – privacy concerns, low-lighting conditions, requirement of detecting micro-motions, etc. This dissertation addresses a set of fundamental and system-level challenges that hinder the wide adoption of mmWave signals in perception systems: increased and rapid temperature increase due to higher power consumption, information loss resulting from specular reflections, networkingsensing performance trade-offs in integrated system, and motion errors and sparse measurements in hand-held operation.

In the final work of this dissertation, we address the limitation of mmWave system, or any singlemodality perception system, to address perception in more complex digital health applications such as the post-stroke recovery assessment. While mmWave signals capture the fine-grained limb motion abnormality and pose asymmetry in stroke survivors, post-stroke symptoms span across a wider range in terms of scale and functionality: from broad body pose and motor dysfunction to fine-grained limb coordination and facial weakness, to speech and conversational impairments. This dissertation presents a first step towards a multimodal approach to automating post-stroke recovery assessment using vision and audio while retaining mmWave signals as a complementary integration. This work is an interdisciplinary collaboration between our lab and researchers from the University of South Carolina, School of Medicine, paving the way towards research in multimodal digital health sensing on a wider scale.

DISSERTATION DEFENSE
 

Author : Ahmed Shehab Khan
Advisors: Dr. Yan Tong
Date: June 22, 2026
Time: 02:00 pm
Place: Virtual (Zoom)
Link:  https://sc-edu.zoom.us/j/86147086296

Abstract
Group Emotion Recognition (GER) is the task of inferring the collective emotional state of a group of individuals from a single image. The task has several inherent challenges. First, relevant evidence is spread across faces, body language, objects, and the surrounding scene, and no single cue is sufficient. Second, individuals and regions within a group contribute unequally to the perceived emotion, with the relative importance shifting from image to image. Third, existing methods that integrate these signals rely on multi-stream pipelines with separate detectors and networks, making inference computationally expensive. This dissertation develops three deep learning frameworks for GER, each building on the previous to address these challenges.

First, we proposed a four-stream hybrid network that combines features from individual faces, the scene, and the spatial arrangement of faces within the image. A face-location aware stream captures the relationship between faces and scene through an attention heatmap; a multi-scale face stream handles the high variance in face size found in images collected in the wild; and a global blurred stream learns scene-only features by suppressing face appearance. These four streams are combined with hand-engineered fusion weights.

Second, we proposed Regional Attention Networks with Context-aware Fusion. Building on the multi-stream approach, this work addressed two limitations: how to determine the importance of individual persons and objects within a group, and how the relative weight of different streams should depend on the image. A regional attention mechanism estimates the importance of each person or object from the image, and a context-aware fusion module replaces the fixed stream weights of the prior framework with values derived from the image content itself. To reduce computational cost, feature extraction is consolidated onto a single shared backbone.

Third, we proposed LG-GER, a language-guided distillation framework that addresses the inference cost of detector-driven multi-stream pipelines. To overcome the lack of spatial supervision in existing GER datasets, a multimodal large language model serves as an offline annotator that generates dense, spatially grounded emotion evidence (bounding boxes with emotion signals and confidence scores) for the training images. This structured evidence is distilled into a single vision-language backbone through four complementary losses: classification, region-text grounding, spatial emotion, and spatial confidence regression. At inference, the framework requires no detectors, no MLLMs, and no multi-stream fusion, making it the first detector-free framework for group emotion recognition.

All three frameworks were evaluated on publicly available GER benchmarks. Visualization and case studies illustrate how the attention and fusion components identify the most informative regions for group emotion recognition.

DISSERTATION DEFENSE

Author : Md Abir Hossen
Advisors: Dr. Pooyan Jamshidi
Date: May 22, 2026
Time: 10:30 am
Place: Virtual (Zoom)
Link:  https://sc-edu.zoom.us/j/88165893836?pwd=wVzSgMF42SJFqBHkE7QtSQO1Tsemp4…

Abstract

Cyber-physical robotic systems expose a combinatorially large configuration space comprising interacting hardware and software parameters. Incorrect configurations can lead to functional faults that are difficult to diagnose due to the intricate and often hidden dependencies between system settings and performance. This dissertation addresses these challenges by learning and exploiting causal structure to enable robust, sample-efficient, and transferable configuration optimization.

The first part of this work introduces CaRE (Causal Robotics DEbugging), a causal diagnosis framework that identifies the root causes of observed functional faults. By learning causal relationships between configuration parameters and performance indicators from observational data, CaRE enables precise fault localization and validation through targeted interventions across both simulation and physical robot platforms.

Building on this causal foundation, the next stage introduces CURE (Causal Understanding and Remediation for Enhancing Robot Performance), a configuration optimization method that identifies causally relevant parameters and restricts optimization to a reduced subspace. CURE improves convergence efficiency and supports transfer across environments by leveraging causal knowledge obtained in low-cost simulations and applying it to real-world robot deployments.

The final part of this dissertation introduces RESCUE (REducing  Sampling cost with Causal Understanding and Estimation), which extends the optimization setting from a single-fidelity target environment to multi-fidelity settings where multiple information sources with different costs and accuracies are available. RESCUE uses causal structure to construct a causal prior and guide configuration-fidelity selection, reducing costly high-fidelity evaluations while preserving optimization quality. Empirical evaluation across synthetic and real-world problems shows improved sample efficiency, more effective fidelity allocation, and lower constraint violation rates than competing methods.

Collectively, this dissertation establishes a causal foundation for reliable, efficient, and transferable configuration debugging and optimization in cyber-physical systems, validated through both synthetic benchmarks and real-world robotic applications.
 

  DISSERTATION DEFENSE

Author : Lai Wei
Advisors: Dr. Jianjun Hu
Date: May 19, 2026
Time: 10:00 am
Place: Room 2267
Virtual (Zoom)
Link: https://sc-edu.zoom.us/j/4997546955#success

Abstract

The discovery of novel inorganic materials is fundamental to technological progress, yet traditional experimental approaches are often slow and resource-intensive. While computational methods have emerged to accelerate this process, significant bottlenecks remain in both the generative design of new chemical compositions and the rapid, accurate prediction of their crystal structures. Furthermore, the field of Crystal Structure Prediction (CSP) has historically lacked a standardized, quantitative framework for evaluating and comparing the performance of diverse algorithms, hindering reproducible progress. Finally, the rapid proliferation of deep generative models for CSP has introduced a new challenge: the absence of a reproducible, leakage-controlled evaluation protocol tailored to this emerging paradigm.

This dissertation presents a comprehensive, end-to-end methodology that addresses these interconnected challenges. The research narrative begins with the development of the Blank-filling Language Model for Materials (BLMM), a deep learning language model that learns the "chemical grammar" of materials to generate novel, chemically plausible compositions. To bridge the gap between composition and structure, we then introduce the Template-Based Crystal Structure Prediction (TCSP) algorithm, a high-throughput method for rapidly predicting structures for these newly generated formulas.

In the process of developing these tools, we identified the critical need for a universal evaluation framework. To this end, we propose CSPMetrics, a systematic suite of quantitative metrics, and CSPBench, a benchmark platform for the fair and rigorous assessment of a broad range of CSP algorithms. Leveraging insights gained directly from CSPBench, which highlighted the efficacy of template-based approaches, we developed TCSP 2.0, a significantly improved algorithm incorporating superior oxidation state prediction and refined chemical heuristics.

Building on this evaluation framework, we further conduct a standardized, leakage-controlled assessment of twelve representative deep generative CSP algorithms, spanning latent-variable, diffusion-based, flow-based, and autoregressive architectures. We introduce a rigorously filtered CLEAN test subset to assess true generalization capability, and perform an ablation study revealing that current diffusion-based models remain substantially dependent on structural prototype coverage in their training data.

This work establishes a complete research cycle: we generate, we predict, we establish a standard to evaluate, we use that standard to innovate and improve, and we apply that standard to comprehensively assess the next generation of deep learning methods. Together, these contributions provide the materials science community with an integrated, data-driven framework to accelerate the discovery of next-generation functional materials.

DISSERTATION DEFENSE

Author : Hong Yung (Joey) Yip
Advisors: Dr. Amit Sheth
Date: April 20th, 2026
Time: 9:30 am
Place: 529 Seminar Room, AI Institute
Link:  https://sc-edu.zoom.us/j/8440139296?omn=81104593412

Abstract

Modern Artificial Intelligence (AI) systems, particularly Large Language Models (LLMs), are powerful but remain largely black-box systems whose outputs often lack explicit semantics, provenance, and temporal validity. In high-stakes enterprise and regulatory settings, these limitations reduce trust, auditability, and reliability. This dissertation argues that data alone is not enough and presents a Neuro-symbolic AI framework (EMPWR) for the Knowledge Graph (KG) lifecycle that integrates symbolic knowledge representation with data-driven learning to improve interoperability, traceability, and grounded response quality.

The dissertation makes four primary contributions, each aligned with a phase of the KG lifecycle. First, for data interoperability, it introduces an ontology-driven framework for modeling non-sequential processes and integrating fragmented data through a Sales Engagement Graph (SEG). Second, for knowledge representation, it introduces the Dynamic-Singleton Property Graph (D-SPG) which combines the semantic rigor of Resource Description Framework (RDF) with the efficiency of Property Graphs (PGs) while modeling provenance, attribution, and temporal validity as first-class elements. The model preserves SPARQL-compliant querying while enabling auditable, metadata-rich knowledge representation.

Third, for alignment and enrichment, the dissertation proposes Context-Enriched Learning Models that leverage hierarchical graph structure, source synonyms, and semantic groups in the UMLS Metathesaurus to improve biomedical vocabulary alignment. Fourth, for consumption and evaluation, it introduces VERIFY, the Validated Evidence Retrieval & Integrity Framework for High-Fidelity Medical Device Question Answering over FDA medical device records. VERIFY contributes a retrieval-aware metric, Response Fidelity (RF), which measures LLM response on correctness, grounding, omission, evidence integrity, and structural validity. These contributions establish EMPWR as an end-to-end framework for building, maintaining, and evaluating trustworthy KG-grounded AI systems, demonstrating that explicit domain knowledge and principled knowledge representation are essential for reliable AI in high-stakes settings.

Abstract:

Analyzing non-stationary signals remains a key challenge across many areas of science and engineering, from biomedical measurements to environmental and mechanical systems. Traditional Fourier-based methods often fail to capture transient or evolving patterns that carry critical information. In this talk, I will introduce time-frequency distributions as a versatile framework for examining complex, dynamic signals. I will also discuss compressive sensing, an advanced technique that enables accurate reconstruction of signals from limited measurements, improving efficiency in data acquisition and analysis. When combined with machine learning methods, these approaches offer powerful tools for feature extraction, pattern recognition, and interpretation across diverse signal types. Through case studies including EEG recordings and tire sensor data as representative real-world examples, I will illustrate how these techniques address practical challenges in signal analysis.

Biography:

Vedran Jurdana received his Ph.D. in Electrical Engineering from the Faculty of Engineering, University of Rijeka, Croatia, in 2023. He is a Postdoctoral Researcher at the Department of Automation and Electronics, University of Rijeka, where he leads the Laboratory for Electronics. His research focuses on EEG signal processing, machine learning, compressive sensing, eye-tracking, and assistive technologies, with an emphasis on data-driven approaches for biomedical signal analysis. He has led and contributed to multiple scientific research projects funded by the Croatian Science Foundation and the University of Rijeka. Dr. Jurdana has established international research collaborations through visits to Graz University of Technology and Johannes Kepler University Linz (Austria), ELTE University in Budapest (Hungary), and Tampere University (Finland). He has served as a member of the Technical Program Committees of international conferences, including CoBCom 2024 and ATAAC 2026. He delivered an invited talk at the ASPAI 2025 conference and has presented his research at international institutions during his research visits. In recognition of his scientific contributions, he received the University of Rijeka Foundation Award for 2024 in the category of Young Scientist in Technical and Biotechnical Sciences. He has been actively involved in higher education teaching since 2017 and has authored numerous scientific publications in signal processing and biomedical engineering.

Author : Vishal Pallagani
Advisor: Dr. Biplav Srivastava
Date: March 26th, 2026
Time: 9 am
Place: Seminar Room, AI Institute
Link: https://teams.microsoft.com/l/meetup-join/19%3ameeting_NzkzZTk0NGItOGQx…

Abstract

Planning is a fundamental capability for intelligent systems, yet classical planners remain difficult to scale and generalize across domains due to their reliance on hand-engineered models, brittle search procedures, and limited adaptability. These limitations constrain real-world applications that require flexible reasoning, robust decision-making, and the ability to operate across diverse environments.

Large language models have emerged as powerful learners supported by their architectures and large-scale training data, demonstrating strong capabilities in tasks such as question answering, code generation, summarization, and reasoning. Motivated by their broad generalization abilities, I investigate whether language models can be systematically leveraged to advance automated planning.

I begin by conducting a comprehensive study of existing literature to categorize how language models are being used for automated planning, resulting in a structured taxonomy that captures the full landscape of techniques, objectives, and use cases, along with a semi-automated platform for tracking the evolution of this rapidly growing area. Building on this foundation, I examine how language models can be used for effective plan generation by evaluating their pretrained capabilities and quantifying the benefits of fine-tuning on planning-specific data.

To alleviate the limitations of relying solely on language models, I introduce a complementary approach for plan generation by training a compact foundational model from scratch. This approach adopts a state-centric perspective to generalized planning, learning transition dynamics over graph-based state representations. In addition, I develop a planning ontology to systematically capture metadata, enabling more structured data representation and facilitating improved training of future models. I also introduce plan summarization as a downstream task, demonstrating how these trained models can be effectively leveraged for concise and structured plan understanding.

Next, I develop neurosymbolic architectures, grounded in the SOFAI ``Thinking, Fast and Slow'' paradigm, that integrate language models with symbolic planning under metacognitive control to achieve more robust and generalizable plan generation. These architectures address key limitations of language models in planning and demonstrate how symbolic reasoning can complement learning-based models. Finally, I evaluate the resulting generalized planners in real-world applications including dialog-based information retrieval, trustworthy conversational AI for sensitive domains, and adaptive replanning in stochastic manufacturing environments, showing how generalized planners combining language models and symbolic methods are practical tools for deployment.

Collectively, my contributions advance the scientific understanding of how language models can support, improve, and generalize automated planning, offering a coherent path toward planning systems that combine the strengths of learning and symbolic reasoning.

 DISSERTATION DEFENSE

Author : Ningqiao Li
Advisor: Dr. Yan Tong
Date: March 24th, 2026
Time: 3:30 pm
Place: Virtual
Link:  https://teams.microsoft.com/meet/27045487833541?p=PtaQgpwzZvxJkYDZYu
Meeting ID: 270 454 878 335 41
Passcode: TP2Wm3dw
Abstract

As images have become a critical strategy for hospitality businesses to position and differentiate their brands online, and brand personality projected by brands and perceived by prospective consumers has been recognized as a key factor influencing booking decisions, this thesis investigates an understudied area: how brand personality perceived from hotel images can be automatically assessed using advanced computational models.

Specifically, this study systematically compares four model architectures (i.e., YOLO26 Nano, ResNet50, Swin-Small Transformer, and CLIP) across multiple label sources, including human rater annotations, labels generated by a single large language model (LLM) (GPT-4o), and average labels generated by multiple LLMs (GPT-4o, Gemini 2.5 Flash, and Claude Sonnet). A dataset of 2,182 hotel-generated images posted on a social media platform was annotated and evaluated across six brand personality dimensions: relaxing, hospitable, lively, distinctive, sophisticated, and wholesome.
The results demonstrate that CLIP, trained on multi-LLM averaged labels, achieves the highest performance, outperforming all image-only architectures as well as models trained on human annotations or a single LLM. This study contributes to a better understanding of how affective semantics can be effectively recognized by comparing different deep learning models and examining performance differences between models trained on human-labeled data and those trained on generative AI–labeled data. It further extends the discussion on the effectiveness of LLM-generated labels in contexts that require domain knowledge and higher-level semantic interpretation.