{"id":{"repo_id":"msu","oai_identifier":"oai:d.lib.msu.edu:etd_53063"},"canonical_url":"https://search.ndltd.org/etd/msu/oai:d.lib.msu.edu:etd_53063","repository":{"repo_id":"msu","name":"Michigan State University","base_url":"https://d.lib.msu.edu/oai"},"display":{"title":"Toward Efficient and Secure Machine Learning","abstract":"Thesis (Ph.D.)--Michigan State University. Computer Science - Doctor of Philosophy, 2026","abstract_html":"Thesis (Ph.D.)--Michigan State University. Computer Science - Doctor of Philosophy, 2026","abstract_has_math":false,"creators":["Ao, Wei"],"institution":null,"degree_name":null,"degree_level":null,"degree_discipline":null,"degree_department":null,"school":null,"contributors":["Boddeti, Vishnu","Yan, Qiben","Kong, Yu","Zhu, Zhuangdi"],"advisors":[],"committee_chairs":[],"committee_members":[],"year":2026,"date_issued":"2026","date_published":"2026","updated_at":"2026-11-18T18:59:37Z","subjects":["Computer science"],"languages":["eng"],"rights":["Attribution-NonCommercial-ShareAlike 4.0 International"],"rights_urls":[],"identifier_entries":[{"key":"dc:identifier","label":"Identifier","values":["etd:53063","local:Ao_grad.msu_0128D_22072"],"render_values":[{"text":"etd:53063","href":null,"code":true},{"text":"local:Ao_grad.msu_0128D_22072","href":null,"code":true}]}],"person_identifier_entries":[]},"links":{"outbound_url":"https://doi.org/doi:10.25335/xs91-vj22","outbound_label":"DOI","outbound_source":"dc:identifier","record_url":"https://doi.org/doi:10.25335/xs91-vj22","record_label":"Repository record","record_kind":"doi","record_source":"dc:identifier","download_url":null,"download_label":null,"download_source":null,"access_status":null,"embargo_until":null,"access_note":null},"source_record":{"url":"https://d.lib.msu.edu/oai?verb=GetRecord&metadataPrefix=oai_dc&identifier=oai%3Ad.lib.msu.edu%3Aetd_53063","prefix":"oai_dc"},"metadata_groups":[{"id":"people","label":"People","entries":[{"key":"dc:contributor","label":"Contributor","values":["Boddeti, Vishnu","Yan, Qiben","Kong, Yu","Zhu, Zhuangdi"]},{"key":"dc:creator","label":"Author","values":["Ao, Wei"]}]},{"id":"academic_context","label":"Academic Context","entries":[{"key":"dc:date","label":"Dc Date","values":["2026"]},{"key":"dc:relation","label":"Dc Relation","values":["Electronic Theses & Dissertations"]},{"key":"dc:type","label":"Dc Type","values":["text","Electronic dissertations"]}]},{"id":"subjects_keywords","label":"Subjects and Keywords","entries":[{"key":"dc:subject","label":"Dc Subject","values":["Computer science"]}]},{"id":"language_rights","label":"Language and Rights","entries":[{"key":"dc:language","label":"Dc Language","values":["eng"]},{"key":"dc:rights","label":"Dc Rights","values":["Attribution-NonCommercial-ShareAlike 4.0 International"]}]},{"id":"identifiers","label":"Identifiers","entries":[{"key":"dc:identifier","label":"Identifier","values":["etd:53063","local:Ao_grad.msu_0128D_22072","https://doi.org/doi:10.25335/xs91-vj22"]}]},{"id":"additional","label":"Additional Metadata","entries":[{"key":"dc:description","label":"Description","values":["Thesis (Ph.D.)--Michigan State University. Computer Science - Doctor of Philosophy, 2026","With the widespread adoption of neural networks, machine learning (ML) has been integrated into our daily lives as a new type of cloud computing service, namely machine learning as a service (MLaaS). Novel ML services have emerged thanks to advances in deep learning, ranging from image recognition and face recognition to conversational and multimodal applications. At the same time, there have emerged privacy and efficiency challenges. In MLaaS, cloud servers provide ML services to clients, who have to upload their sensitive or confidential data, such as biometric data, financial data, and medical data, to these servers. Cryptographic techniques, like fully homomorphic encryption (FHE), provide robust security guarantees for private data but significantly increase inference latency and computational complexity. Furthermore, recent multimodal applications, like video understanding, impose greater challenges on efficiency, especially as video duration increases. First, AutoFHE is introduced, which is an automated machine learning (AutoML) framework to adapt standard neural networks to FHENets for efficient evaluation on FHE. Fully homomorphic encryption (FHE) allows computation on encrypted data without decrypting it. Secure inference of machine learning with FHE opens a new avenue toward secure machine learning (SecureML). However, SecureML is about 3--6 orders of magnitude slower, which poses a significant efficiency challenge. AutoFHE searches for layerwise mixed-degree polynomial activation functions, which are optimized jointly with the homomorphic evaluation architecture in terms of the placement of bootstrapping operations. AutoFHE evolves efficient layerwise polynomial activations with smaller multiplicative depth and reduces expensive bootstrapping operations. The problem is modeled within a multi-objective optimization framework to maximize accuracy and minimize the number of bootstrapping operations. AutoFHE can be flexibly applied to any neural architectures, and it provides diverse solutions that span the trade-off between accuracy and latency. Then, FHENets are applied to a security-critical application, namely face recognition. Face recognition is central to many authentication, security, and personalized applications. Yet, face recognition suffers from significant privacy risks, particularly arising from unauthorized access to sensitive biometric data. Prior secure face recognition with FHE only protects templates and suffers from template recovery attacks. To address these security vulnerabilities, CryptoFace is proposed, which is the first end-to-end encrypted face recognition system with FHE. CryptoFace enables secure processing of facial data across all stages of a face-recognition process—feature extraction, storage, and matching—without exposing raw images or features. To efficiently evaluate high-resolution face images, CryptoFaceNets leverage a mixture of efficient CNNs with small multiplicative depth and only use one bootstrapping operation. CryptoFace will enable secure face recognition systems requiring robust and provable security. Finally, the applications are extended from unimodal to multimodal domains to address the accuracy and efficiency issues in video understanding. Recently, multimodal models, specifically large vision-language models (VLMs), have demonstrated significant advancements in vision-language understanding and reasoning capabilities because of the scaling ability of large language models (LLMs). However, VLMs exhibit performance degradation on (extremely) long videos. Naively sampling more frames does not necessarily improve performance and results in low efficiency and high GPU memory usage. To address these issues, a novel query-to-frame multimodal retrieval, referred to as QSVideo, is proposed to discover supportive visual evidence by jointly optimizing semantic relevance, visual diversity and temporal coverage. QSVideo is seamlessly integrated into any video VLM and greatly boosts VLMs' performance on challenging long videos and streaming videos under a very limited frame budget. The core of this dissertation lies in efficiency and privacy issues in MLaaS. AutoFHE addresses the most critical problem in FHENets, accelerating inference, and proposes an automated search algorithm to achieve it. To make SecureML more practical for critical real-world applications, SecureML is applied to image classification and face recognition. CryptoFace is an efficient end-to-end secure face recognition system. Finally, QSVideo is designed to facilitate VLMs on long and streaming videos by locating supportive moments in the videos.","Description based on online resource. Title from PDF t.p. (Michigan State University Fedora Repository, viewed ).","Includes bibliographical references."]},{"key":"dc:format","label":"Dc Format","values":["application/pdf","1 online resource (112 p.) : ill."]},{"key":"dc:title","label":"Title","values":["Toward Efficient and Secure Machine Learning"]}]}],"canonical_facts":{"dc:contributor":["Boddeti, Vishnu","Yan, Qiben","Kong, Yu","Zhu, Zhuangdi"],"dc:creator":["Ao, Wei"],"dc:date":["2026"],"dc:description":["Thesis (Ph.D.)--Michigan State University. Computer Science - Doctor of Philosophy, 2026","With the widespread adoption of neural networks, machine learning (ML) has been integrated into our daily lives as a new type of cloud computing service, namely machine learning as a service (MLaaS). Novel ML services have emerged thanks to advances in deep learning, ranging from image recognition and face recognition to conversational and multimodal applications. At the same time, there have emerged privacy and efficiency challenges. In MLaaS, cloud servers provide ML services to clients, who have to upload their sensitive or confidential data, such as biometric data, financial data, and medical data, to these servers. Cryptographic techniques, like fully homomorphic encryption (FHE), provide robust security guarantees for private data but significantly increase inference latency and computational complexity. Furthermore, recent multimodal applications, like video understanding, impose greater challenges on efficiency, especially as video duration increases. First, AutoFHE is introduced, which is an automated machine learning (AutoML) framework to adapt standard neural networks to FHENets for efficient evaluation on FHE. Fully homomorphic encryption (FHE) allows computation on encrypted data without decrypting it. Secure inference of machine learning with FHE opens a new avenue toward secure machine learning (SecureML). However, SecureML is about 3--6 orders of magnitude slower, which poses a significant efficiency challenge. AutoFHE searches for layerwise mixed-degree polynomial activation functions, which are optimized jointly with the homomorphic evaluation architecture in terms of the placement of bootstrapping operations. AutoFHE evolves efficient layerwise polynomial activations with smaller multiplicative depth and reduces expensive bootstrapping operations. The problem is modeled within a multi-objective optimization framework to maximize accuracy and minimize the number of bootstrapping operations. AutoFHE can be flexibly applied to any neural architectures, and it provides diverse solutions that span the trade-off between accuracy and latency. Then, FHENets are applied to a security-critical application, namely face recognition. Face recognition is central to many authentication, security, and personalized applications. Yet, face recognition suffers from significant privacy risks, particularly arising from unauthorized access to sensitive biometric data. Prior secure face recognition with FHE only protects templates and suffers from template recovery attacks. To address these security vulnerabilities, CryptoFace is proposed, which is the first end-to-end encrypted face recognition system with FHE. CryptoFace enables secure processing of facial data across all stages of a face-recognition process—feature extraction, storage, and matching—without exposing raw images or features. To efficiently evaluate high-resolution face images, CryptoFaceNets leverage a mixture of efficient CNNs with small multiplicative depth and only use one bootstrapping operation. CryptoFace will enable secure face recognition systems requiring robust and provable security. Finally, the applications are extended from unimodal to multimodal domains to address the accuracy and efficiency issues in video understanding. Recently, multimodal models, specifically large vision-language models (VLMs), have demonstrated significant advancements in vision-language understanding and reasoning capabilities because of the scaling ability of large language models (LLMs). However, VLMs exhibit performance degradation on (extremely) long videos. Naively sampling more frames does not necessarily improve performance and results in low efficiency and high GPU memory usage. To address these issues, a novel query-to-frame multimodal retrieval, referred to as QSVideo, is proposed to discover supportive visual evidence by jointly optimizing semantic relevance, visual diversity and temporal coverage. QSVideo is seamlessly integrated into any video VLM and greatly boosts VLMs' performance on challenging long videos and streaming videos under a very limited frame budget. The core of this dissertation lies in efficiency and privacy issues in MLaaS. AutoFHE addresses the most critical problem in FHENets, accelerating inference, and proposes an automated search algorithm to achieve it. To make SecureML more practical for critical real-world applications, SecureML is applied to image classification and face recognition. CryptoFace is an efficient end-to-end secure face recognition system. Finally, QSVideo is designed to facilitate VLMs on long and streaming videos by locating supportive moments in the videos.","Description based on online resource. Title from PDF t.p. (Michigan State University Fedora Repository, viewed ).","Includes bibliographical references."],"dc:format":["application/pdf","1 online resource (112 p.) : ill."],"dc:identifier":["etd:53063","local:Ao_grad.msu_0128D_22072","https://doi.org/doi:10.25335/xs91-vj22"],"dc:language":["eng"],"dc:relation":["Electronic Theses & Dissertations"],"dc:rights":["Attribution-NonCommercial-ShareAlike 4.0 International"],"dc:subject":["Computer science"],"dc:title":["Toward Efficient and Secure Machine Learning"],"dc:type":["text","Electronic dissertations"]},"updated_at":"2026-11-18T18:59:37Z"}