Computer Science electronic theses and dissertations (MU)
Permanent URI for this collection
The items in this collection are the theses and dissertations written by students of the Department of Computer Science. Some items may be viewed only by members of the University of Missouri System and/or University of Missouri-Columbia. Click on one of the browse buttons above for a complete listing of the works.
Browse
Recent Submissions
Item Robust speech recognition : addressing data scarcity and distant speech with novel modeling and enhancement techniques(University of Missouri--Columbia, 2025) Zhao, Tuo; Zhao, Yunxin[EMBARGOED UNTIL 08/01/2026] In speech recognition, to combat model overfitting and maintain generalization ability in insufficient data conditions, ensemble acoustic modeling that aggregates outputs from multiple sub-models is appreciated. In acoustic modeling (AM), phonetic decision trees (PDTs) are well-known for overcoming data insufficiency and sparsity by clustering context-dependent tri-phone units. Random forest of PDTs has shown a strong ability in further improving Gaussian mixture model-hidden Markov model (HMM) based AM, by generating random forests of PDTs (RF-PDTs) to form RF-PDT-based ensemble acoustic models (EAMs). We thus propose to transfer this concept to context-dependent deep neural network-hidden Markov model (CD-DNNHMM) acoustic modeling scheme, resulting in a novel DNN-HMM EAM approach that utilizes an RF-based ensemble of DNNs. Evaluation results on TIMIT dataset and a telemedicine automatic captioning dataset have demonstrated the superior performance of our proposed RF-PDT+CD-DNN-based EAM over the conventional CD-DNN-based single acoustic model in phone and word recognition accuracies. Convolutional neural networks (CNNs) have shown impressive capacity in acoustic feature learning than fully connected NNs, attributed partially to its higher robustness to speech rate variations or better handling of syllable-length speech context through time-domain convolution, or its higher robustness to speaker-induced frequency shifts or better handling of high-resolution spectrum through frequencydomain convolution. We thus propose a novel acoustic modeling approach by using 2D CNNs to combine the advantages described above for robust speech recognition. Based on earlier findings that recognition performance of acoustic modeling in noisy or reverberant conditions could also benefit from combining information in multiple time context sizes or frequency resolutions, we further investigate the potential of 2D CNN-based EAMs by integrating information over different time spans and/or spectral resolutions. Experimental results on speaker-independent phone recognition tasks of TIMIT and FFMTIMIT have demonstrated that our proposed 2D CNN provided consistent phone error reductions over frequency-domain CNN and DNN for both TIMIT and FFMTIMIT, with more benefits shown for recognizing noisy speech by using clean speech models. Further performance improvements have been observed by EAMs that integrate information over different time spans and/or spectral resolutions. The ability of CNN on modeling temporal and spectral local correlations has been further exploited in acoustic feature learning in noisy and reverberant conditions, in which reverberation and noise often manifested as blurs on both axes in 2D spectrogram images. For this purpose, we propose to utilize a fully convolutional network (FCN) architecture in computer vision, i.e., UNet++, to perform multi-channel speech enhancement for distant speech recognition (DSR), exploiting the robustness of the FCN as well as utilizing the spatial information in multi-channel recordings. DSR results from the multiple distant microphone (MDM) datasets of AMI meeting corpus have demonstrated that our proposed UNet++-based multi-channel speech enhancement approaches provided large word error rate (WER) reductions over its UNet- and weighted prediction error (WPE)-based counterparts, and the UNet++ with WPE preprocessing achieved the lowest WERs. To further advance robustness while alleviating data insufficiency in distant speech recognition (DSR) under multi-channel, reverberant, non-native, and spontaneous multi-speaker conditions, we incorporate a self-supervised learning representation (SSLR)-based acoustic model with a dual-stage UNet++ enhancement frontend and clean-reference augmentation. The SSLR model, built with minimal architectural modification to the existing transformer-based ASR, addresses the challenge of laxiv beled data scarcity by leveraging pre-trained acoustic representations from largescale unlabeled audio, leading to notable WER improvements over the prior baseline. We then reintroduce the multi-channel UNet++ speech enhancement (SE) frontend and demonstrate its effectiveness when combined with the SSLR-based ASR model, yielding further WER reduction. Building on this, we propose a second UNet++ module--positioned after the first-stage regression-based SE and before the ASR backbone--to form an ASR-performance-driven SE layer, bringing additional improvements. We also explore joint optimization of the secondary SE module with the ASR model and introduce a clean-reference-interleaved augmentation strategy that supplies paired clean speech as auxiliary input, further enhancing recognition accuracy. All experiments were conducted using the multiple distant microphone (MDM) recordings from the AMI meeting corpus, ensuring evaluation under realistic DSR conditions. Together, these extensions form a tri-stage SE1-SE2-ASR pipeline that significantly improves recognition performance in low-resource and reverberant speech scenarios.Item Intelligent dialog agent modeling in human-centered artificial intelligence applications(University of Missouri--Columbia, 2025) Oruche, Roland; Prasad, Calyam[EMBARGOED UNTIL 08/01/2026] Dialog systems have been integral in which ways humans and AI models can interact across multiple domains such as healthcare, customer service, and education. With the recent advancements in deep learning and pre-trained language models (PLMs), dialog systems have shown impressive performance in various downstream natural language processing tasks for the purposes of providing confidence in human decision making in real-world applications. Despite the recent success, traditional dialog systems are challenged in two ways: (a) dialog systems that trained/pre-trained autonomously -- which does not require human involvement -- can lead to training inefficiencies, lack of human alignment, and lack of feedback loops; (b) fully autonomous systems that lack prediction justifications can be harmful if humans blindly follow the predictions outcomes. Thus, as humans continue to adopt dialog systems for their domain-specific or personal needs, synergistic efforts between dialog systems and humans are needed to increase trust, transparency, and user satisfaction in realistic applications. In addressing these major challenges, this thesis introduces the design and modeling of "Human- Centered Dialog Systems" (HCDS), which leverages human-centered AI (HCAI) techniques to integrate humans in every facet of the dialog development life cycle. Specifically, we present three novelty areas in which HCDS can be employed to alleviate the aforementioned challenges: (i) Holistic Multi-Layer Dialogs details the design of dialog systems in across a multi-layered development lifecycle, which includes the data manage layer, model management layer, and application layer; (ii) Efficient Human-in-the-Loop Dialogs aims to eliminate the inefficiencies of dialog training by incorporating humans for annotation and labeling feedback; (iii) Trustworthy and Factual Dialogs leverages human teaching and instruction to ensure deployed models for inference in application settings provide transparent and factual answers.Item Active cyber defense using intelligent agents in cloud-edge based applications(University of Missouri--Columbia, 2025) Neupane, Roshan Lal; Calyam, Prasad[EMBARGOED UNTIL 08/01/2026] The need for active cyber defense advances is becoming increasingly apparent, given the sophisticated nature of modern cyber threats. Especially, cloud-edge systems in critical infrastructure domains, such as e.g., healthcare, finance, and smart grids, need to be designed with an adversarial mindset as part of active cyber defense, where knowledge of potential attacks is applied to outsmart the adversaries. While passive cyber defense measures are important, new paradigms for active cyber defense are emerging. This dissertation investigates advanced active cyber defense strategies for cloud-edge infrastructures essential to these domains, addressing the complex and evolving nature of cyber threats targeting these critical domains. The thesis develops proactive and reactive defense methods, exploring three foundational thrusts: (1) game-theoretic deception techniques to mislead and delay adversaries, (2) blockchain-based security mechanisms augmented with formal methods for integrity and auditability, and (3) AI-driven knowledge systems that enable real-time threat detection and mitigation through structured reasoning. First, we develop multistage game-theoretic deception strategies that guide optimal placement of honeypots, honeyfiles, and honeytokens along stages of the ransomware kill chain. Using Subgame-Perfect Nash Equilibrium, the approach optimizes defender actions under adversarial uncertainty. Second, we introduce blockchain infrastructures that ensure tamper-resistance and transparency, incorporate formal verification techniques using Linear Temporal Logic and model checking to validate critical operations, and use ML models for proactive detection of infrastructure- and application-level anomalies. Third, we present knowledge-driven threat modeling, analysis, and mitigation using integrated knowledge graphs (KGs), large language models (LLMs,) and Software Defined Networking (SDN) to enable adaptive threat detection and response mechanisms. These core defense strategies are instantiated and validated in three application domain testbeds. In healthcare, our multistage ransomware defense framework demonstrates the effectiveness of deception-based defense across various attack stages. In the financial sector, we implement ClaimChain, a permissioned consortium blockchain platform that supports secure claims processing through formally verified smart contracts that integrate infrastructure-level attack modeling with applicationlevel fraud detection. In the smart grid domain, we develop (1) CIBR-Fort, a comprehensive cyber defense framework that integrates KG-LLM pipelines for link prediction and reasoning, enabling real-time threat mitigation in smart grid environments, and (2) SGChain, a permissioned blockchain to address availability attacks by securing metering data coupled with a distributed SDN control for network-level attack mitigation.Item Bridging in-game behaviors and learning outcomes : design, implementation, and validation of stealth assessment pipelines in Mission HydroSci(University of Missouri--Columbia, 2025) Lu, Wenyi; Goggins, Sean P.[EMBARGOED UNTIL 08/01/2026] Digital Game-Based Learning (DGBL) holds the potential to foster deep engagement, complex problem-solving, and high motivation among learners. However, fully leveraging these advantages requires robust, scalable, and unobtrusive assessment strategies to capture nuanced cognitive and affective developments in real time. This dissertation addresses these needs through three interlinked studies conducted in the context of Mission HydroSci (MHS), a 3D digital game designed to teach middleschool water science and scientific argumentation skills. Study 1 demonstrates how to design and implement a customized, learning-focused logging system by combining the Activity Theory-based Model of Serious Games (ATMSG) and the Experience API (xAPI). This co-designed approach ensures granular data capture aligned with key educational objectives, effectively distinguishing novice from expert in-game behaviors. Study 2 builds on these logs to create a fully interpretable stealth assessment pipeline for a single learning objective. Guided by conceptual frameworks for systematic feature engineering, the study employs various machine learning classifiers and a surrogate modeling approach to accurately predict targeted competencies, preserving interpretability for educators and stakeholders. Study 3 scales the methodology to multiple learning outcomes across diverse MHS units. By incorporating multi-layered unsupervised learning for feature extraction, ensemble-learning-based predictive modeling, and post-hoc interpretability techniques such as permutation importance scores and Accumulated Local Effects (ALE) plots, this extended pipeline adapts to complex data distributions while maintaining transparency. Findings reveal that purposeful data collection, systematic feature engineering, and robust modeling approaches can reliably capture and forecast learners' knowledge gains, thereby enabling timely feedback and targeted interventions. Collectively, these studies substantiate a coherent progression from conceptual foundations to practical solutions in stealth assessment. The dissertation underscores how embedded logging systems and sophisticated analytics can assess and enhance learning in complex DGBL environments, offering actionable insights for designers, educators, and researchers alike.Item Security and trust in cloud-edge integration for protected data and instrumentation management(University of Missouri--Columbia, 2025) Lemus Alarcon, Mauro Enrique; Calyam, Prasad[EMBARGOED UNTIL 08/01/2026] The convergence of edge-cloud resources--including volunteer computing, remote scientific instrumentation, and protected healthcare data--has become essential for supporting data-intensive workflows, real-time experimentation, and healthcare advances. As scientific applications rely more on distributed, heterogeneous environments, secure and efficient resource integration becomes essential for advancing discovery and innovation. However, this approach faces key challenges: volunteer edge-cloud (VEC) environments--an emerging model for scalable integration--must overcome resource heterogeneity, trust, scalability, and privacy issues; remote instrumentation management struggles with secure connectivity and experiment control; and protected healthcare data collection and brokering are complicated by diverse sources, privacy risks, and manual sharing. These barriers often result in fragmented solutions that limit collaborative, cross-disciplinary research. This thesis addresses these challenges with three key contributions. First, for VEC computing, mechanisms are introduced to foster trust in volunteer environments, machine learning-based schedulers are proposed for optimal and reliable task allocation, scalable resource management strategies are developed, and privacy is enabled for both data and models. Second, to advance remote instrumentation management, the Remote Instrumentation Scientific Environment (RISE) architecture is presented, enabling secure integration of instruments and data storage, streamlining experimental setup and control, and facilitating secure, collaborative sharing of experimental results. Third, for protected healthcare data, a semi-automated honest brokering system is proposed that utilizes a common data model to aggregate data from multiple, diverse sources and applies a trust-based mechanism to expedite and secure healthcare data requests for research. Together, these contributions significantly enhance the reliability, scalability, and security of cloud-edge resource integration, enabling more effective scientific instrumentation and data management.