{"id":{"repo_id":"gmu","oai_identifier":"oai:MARS:1920/14754"},"canonical_url":"https://search.dev.ndltd.org/etd/gmu/oai:MARS:1920/14754","repository":{"repo_id":"gmu","name":"George Mason University","base_url":"https://mars.gmu.edu/server/oai/request"},"display":{"title":"Towards High-Performance and Secure Machine Learning on Advanced Computing Platforms","abstract":"Along with the rapid progress in the algorithms and the underlying computing platforms for machine learning (ML), the standards for the performance and security of ML systems have risen significantly. For the design of ML algorithms, the recent breakthrough in transformer-based pretrained language models (LMs) has shown substantial performance improvements across various natural language processing (NLP) tasks. Their strong ability to understand the context of language sequence also inspires the development of multi-modality LMs for text, images and audio. However, the potential of LMs in graph machine learning has not been fully activated, as graph neural networks (GNNs) remain the state-of-the-art algorithms for a large number of graph-based applications. In addition, unique security challenges are emerging when applying LMs to real-world applications, including evasion attacks like jailbreaking, training data poisoning such as backdoor attacks, prompt injection attacks, and model stealing attacks of closed-source large language models (LLMs). Recently, a new training data extraction attack was developed by leveraging the memorization of LLMs. To defend and mitigate the security concerns incurred by memorization, it is critical to quantify the memorization capability of LLMs accurately. However, the existing evaluation methods tend to underestimate the memorization of a given LLM. In parallel with advancements in the algorithm design, the underlying computing platforms for ML are also rapidly evolving, with quantum computing emerging as one of the most promising platforms. More specifically, the unique advantages of quantum computing, coupled with the continuous and fast progress in constructing actual quantum computers, have given rise to the research on quantum machine learning (QML), which aims to implement and accelerate ML algorithms on quantum computers. However, there still exists a big performance gap between the classic ML algorithms and the existing QML algorithms. In addition, current access to actual quantum computers for ordinary users is limited to remote usage through cloud services, named Quantum-as-a-Service (QaaS). Therefore, once the QML algorithms are deployed on quantum computers, the QML services have to be provided to ordinary users via the QaaS paradigm, which poses high security risks due to the leakage of well-tuned QML models and results in substantial monetary loss for the model developer. To solve the above challenges, the dissertation proposes (1). A self-guided multimodal approach to enhance domain-specific graph machine learning, which leverages the power of pretrained LM and is effective in the diagnosis of Alzheimer’s Disease; (2). A new method to estimate the LLM's memorization via utilizing the dynamic soft prompting generated by a trained generator to react to the change of input and thus enabling more accurate extraction of memorized data; (3). QF-Mixer, a framework to design a quantum neural architecture by mixing different types of quantum neurons, which outperforms the existing QML algorithms by a large margin; (4). QuMoS, a framework for preserving the security of QML models via distributed inference without hurting the model performance.","abstract_html":"Along with the rapid progress in the algorithms and the underlying computing platforms for machine learning (ML), the standards for the performance and security of ML systems have risen significantly. For the design of ML algorithms, the recent breakthrough in transformer-based pretrained language models (LMs) has shown substantial performance improvements across various natural language processing (NLP) tasks. Their strong ability to understand the context of language sequence also inspires the development of multi-modality LMs for text, images and audio. However, the potential of LMs in graph machine learning has not been fully activated, as graph neural networks (GNNs) remain the state-of-the-art algorithms for a large number of graph-based applications. In addition, unique security challenges are emerging when applying LMs to real-world applications, including evasion attacks like jailbreaking, training data poisoning such as backdoor attacks, prompt injection attacks, and model stealing attacks of closed-source large language models (LLMs). Recently, a new training data extraction attack was developed by leveraging the memorization of LLMs. To defend and mitigate the security concerns incurred by memorization, it is critical to quantify the memorization capability of LLMs accurately. However, the existing evaluation methods tend to underestimate the memorization of a given LLM. In parallel with advancements in the algorithm design, the underlying computing platforms for ML are also rapidly evolving, with quantum computing emerging as one of the most promising platforms. More specifically, the unique advantages of quantum computing, coupled with the continuous and fast progress in constructing actual quantum computers, have given rise to the research on quantum machine learning (QML), which aims to implement and accelerate ML algorithms on quantum computers. However, there still exists a big performance gap between the classic ML algorithms and the existing QML algorithms. In addition, current access to actual quantum computers for ordinary users is limited to remote usage through cloud services, named Quantum-as-a-Service (QaaS). Therefore, once the QML algorithms are deployed on quantum computers, the QML services have to be provided to ordinary users via the QaaS paradigm, which poses high security risks due to the leakage of well-tuned QML models and results in substantial monetary loss for the model developer. To solve the above challenges, the dissertation proposes (1). A self-guided multimodal approach to enhance domain-specific graph machine learning, which leverages the power of pretrained LM and is effective in the diagnosis of Alzheimer’s Disease; (2). A new method to estimate the LLM's memorization via utilizing the dynamic soft prompting generated by a trained generator to react to the change of input and thus enabling more accurate extraction of memorized data; (3). QF-Mixer, a framework to design a quantum neural architecture by mixing different types of quantum neurons, which outperforms the existing QML algorithms by a large margin; (4). QuMoS, a framework for preserving the security of QML models via distributed inference without hurting the model performance.","abstract_has_math":false,"creators":["Wang, Zhepeng"],"institution":null,"degree_name":null,"degree_level":null,"degree_discipline":null,"degree_department":null,"school":null,"contributors":[],"advisors":[],"committee_chairs":[],"committee_members":[],"year":2025,"date_issued":"2025","date_published":"2025","updated_at":"2026-07-27T19:52:04Z","subjects":["Language Model","Machine Learning","Model Security","Natural Language Processing","Quantum Computing","Quantum Machine Learning"],"languages":[],"rights":[],"rights_urls":[],"identifier_entries":[{"key":"dc:identifier","label":"Identifier","values":["hdl:1920/14754"],"render_values":[{"text":"hdl:1920/14754","href":null,"code":true}]}]},"links":{"outbound_url":null,"outbound_label":null,"outbound_source":null},"metadata_groups":[{"id":"academic_context","label":"Academic Context","entries":[{"key":"dc:date.issued","label":"Date","values":["2025"]},{"key":"dc:type","label":"Dc Type","values":["Dissertation"]}]},{"id":"subjects_keywords","label":"Subjects and Keywords","entries":[{"key":"dc:subject","label":"Dc Subject","values":["Language Model","Machine Learning","Model Security","Natural Language Processing","Quantum Computing","Quantum Machine Learning"]}]},{"id":"identifiers","label":"Identifiers","entries":[{"key":"dc:identifier","label":"Identifier","values":["hdl:1920/14754"]}]},{"id":"additional","label":"Additional Metadata","entries":[{"key":"dc:description.other","label":"Dc Description Other","values":["Along with the rapid progress in the algorithms and the underlying computing platforms for machine learning (ML), the standards for the performance and security of ML systems have risen significantly. For the design of ML algorithms, the recent breakthrough in transformer-based pretrained language models (LMs) has shown substantial performance improvements across various natural language processing (NLP) tasks. Their strong ability to understand the context of language sequence also inspires the development of multi-modality LMs for text, images and audio. However, the potential of LMs in graph machine learning has not been fully activated, as graph neural networks (GNNs) remain the state-of-the-art algorithms for a large number of graph-based applications. In addition, unique security challenges are emerging when applying LMs to real-world applications, including evasion attacks like jailbreaking, training data poisoning such as backdoor attacks, prompt injection attacks, and model stealing attacks of closed-source large language models (LLMs). Recently, a new training data extraction attack was developed by leveraging the memorization of LLMs. To defend and mitigate the security concerns incurred by memorization, it is critical to quantify the memorization capability of LLMs accurately. However, the existing evaluation methods tend to underestimate the memorization of a given LLM. In parallel with advancements in the algorithm design, the underlying computing platforms for ML are also rapidly evolving, with quantum computing emerging as one of the most promising platforms. More specifically, the unique advantages of quantum computing, coupled with the continuous and fast progress in constructing actual quantum computers, have given rise to the research on quantum machine learning (QML), which aims to implement and accelerate ML algorithms on quantum computers. However, there still exists a big performance gap between the classic ML algorithms and the existing QML algorithms. In addition, current access to actual quantum computers for ordinary users is limited to remote usage through cloud services, named Quantum-as-a-Service (QaaS). Therefore, once the QML algorithms are deployed on quantum computers, the QML services have to be provided to ordinary users via the QaaS paradigm, which poses high security risks due to the leakage of well-tuned QML models and results in substantial monetary loss for the model developer. To solve the above challenges, the dissertation proposes (1). A self-guided multimodal approach to enhance domain-specific graph machine learning, which leverages the power of pretrained LM and is effective in the diagnosis of Alzheimer’s Disease; (2). A new method to estimate the LLM's memorization via utilizing the dynamic soft prompting generated by a trained generator to react to the change of input and thus enabling more accurate extraction of memorized data; (3). QF-Mixer, a framework to design a quantum neural architecture by mixing different types of quantum neurons, which outperforms the existing QML algorithms by a large margin; (4). QuMoS, a framework for preserving the security of QML models via distributed inference without hurting the model performance."]},{"key":"dc:title","label":"Title","values":["Towards High-Performance and Secure Machine Learning on Advanced Computing Platforms"]}]}],"canonical_facts":{"dc:date.issued":["2025"],"dc:description.other":["Along with the rapid progress in the algorithms and the underlying computing platforms for machine learning (ML), the standards for the performance and security of ML systems have risen significantly. For the design of ML algorithms, the recent breakthrough in transformer-based pretrained language models (LMs) has shown substantial performance improvements across various natural language processing (NLP) tasks. Their strong ability to understand the context of language sequence also inspires the development of multi-modality LMs for text, images and audio. However, the potential of LMs in graph machine learning has not been fully activated, as graph neural networks (GNNs) remain the state-of-the-art algorithms for a large number of graph-based applications. In addition, unique security challenges are emerging when applying LMs to real-world applications, including evasion attacks like jailbreaking, training data poisoning such as backdoor attacks, prompt injection attacks, and model stealing attacks of closed-source large language models (LLMs). Recently, a new training data extraction attack was developed by leveraging the memorization of LLMs. To defend and mitigate the security concerns incurred by memorization, it is critical to quantify the memorization capability of LLMs accurately. However, the existing evaluation methods tend to underestimate the memorization of a given LLM. In parallel with advancements in the algorithm design, the underlying computing platforms for ML are also rapidly evolving, with quantum computing emerging as one of the most promising platforms. More specifically, the unique advantages of quantum computing, coupled with the continuous and fast progress in constructing actual quantum computers, have given rise to the research on quantum machine learning (QML), which aims to implement and accelerate ML algorithms on quantum computers. However, there still exists a big performance gap between the classic ML algorithms and the existing QML algorithms. In addition, current access to actual quantum computers for ordinary users is limited to remote usage through cloud services, named Quantum-as-a-Service (QaaS). Therefore, once the QML algorithms are deployed on quantum computers, the QML services have to be provided to ordinary users via the QaaS paradigm, which poses high security risks due to the leakage of well-tuned QML models and results in substantial monetary loss for the model developer. To solve the above challenges, the dissertation proposes (1). A self-guided multimodal approach to enhance domain-specific graph machine learning, which leverages the power of pretrained LM and is effective in the diagnosis of Alzheimer’s Disease; (2). A new method to estimate the LLM's memorization via utilizing the dynamic soft prompting generated by a trained generator to react to the change of input and thus enabling more accurate extraction of memorized data; (3). QF-Mixer, a framework to design a quantum neural architecture by mixing different types of quantum neurons, which outperforms the existing QML algorithms by a large margin; (4). QuMoS, a framework for preserving the security of QML models via distributed inference without hurting the model performance."],"dc:identifier":["hdl:1920/14754"],"dc:subject":["Language Model","Machine Learning","Model Security","Natural Language Processing","Quantum Computing","Quantum Machine Learning"],"dc:title":["Towards High-Performance and Secure Machine Learning on Advanced Computing Platforms"],"dc:type":["Dissertation"]},"updated_at":"2026-07-27T19:52:04Z"}