{"id":{"repo_id":"msu","oai_identifier":"oai:d.lib.msu.edu:etd_53004"},"canonical_url":"https://search.ndltd.org/etd/msu/oai:d.lib.msu.edu:etd_53004","repository":{"repo_id":"msu","name":"Michigan State University","base_url":"https://d.lib.msu.edu/oai"},"display":{"title":"Data protection in generative AI : from data owners to model builders","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":["Ren, Jie"],"institution":null,"degree_name":null,"degree_level":null,"degree_discipline":null,"degree_department":null,"school":null,"contributors":["Tang, Jiliang","Xing, Yue","Liu, Hui","Tang, Xianfeng"],"advisors":[],"committee_chairs":[],"committee_members":[],"year":2026,"date_issued":"2026","date_published":"2026","updated_at":"2026-11-18T18:59:34Z","subjects":["Artificial intelligence"],"languages":["eng"],"rights":["In Copyright"],"rights_urls":[],"identifier_entries":[{"key":"dc:identifier","label":"Identifier","values":["etd:53004","local:Ren_grad.msu_0128D_22007"],"render_values":[{"text":"etd:53004","href":null,"code":true},{"text":"local:Ren_grad.msu_0128D_22007","href":null,"code":true}]}],"person_identifier_entries":[]},"links":{"outbound_url":"https://doi.org/doi:10.25335/wv4s-kt58","outbound_label":"DOI","outbound_source":"dc:identifier","record_url":"https://doi.org/doi:10.25335/wv4s-kt58","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_53004","prefix":"oai_dc"},"metadata_groups":[{"id":"people","label":"People","entries":[{"key":"dc:contributor","label":"Contributor","values":["Tang, Jiliang","Xing, Yue","Liu, Hui","Tang, Xianfeng"]},{"key":"dc:creator","label":"Author","values":["Ren, Jie"]}]},{"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":["Artificial intelligence"]}]},{"id":"language_rights","label":"Language and Rights","entries":[{"key":"dc:language","label":"Dc Language","values":["eng"]},{"key":"dc:rights","label":"Dc Rights","values":["In Copyright"]}]},{"id":"identifiers","label":"Identifiers","entries":[{"key":"dc:identifier","label":"Identifier","values":["etd:53004","local:Ren_grad.msu_0128D_22007","https://doi.org/doi:10.25335/wv4s-kt58"]}]},{"id":"additional","label":"Additional Metadata","entries":[{"key":"dc:description","label":"Description","values":["Thesis (Ph.D.)--Michigan State University. Computer Science - Doctor of Philosophy, 2026","The rapid advancement of generative models, including text-to-image diffusion models and large language models (LLMs), has transformed content creation across modalities. However, these powerful models raise significant concerns regarding data protection, including copyright infringement, privacy leakage, unauthorized dataset usage, and the generation of harmful content. Addressing these risks requires systematic solutions that bridge the perspectives of both data owners and model builders. This dissertation presents a comprehensive study of data protection in generative AI, spanning memorization analysis, dataset auditing, unlearning enhancement, and robustness against malicious exploitation.First, we investigate memorization in text-to-image diffusion models through the lens of cross-attention mechanisms. We reveal that memorization is closely associated with disproportionate attention concentration on specific trigger tokens, leading to overfitting to training samples. Based on this insight, we propose efficient detection metrics grounded in attention entropy and layer-wise analysis, as well as inference-time and training-time mitigation strategies. Our methods significantly reduce memorization without degrading image quality or increasing computational overhead.Second, we address the challenge of removing copyrighted and privacy-sensitive data from LLMs through unlearning. We conduct a systematic analysis of fine-tuning-based unlearning methods and uncover a shared underlying mechanism between gradient-ascent-based and suppression-based approaches: both operate by distinguishing target data and suppressing related generations rather than truly erasing knowledge. Motivated by this finding, we propose Gated Representation UNlearning (GRUN), a general and parameter-efficient framework that combines a soft gating mechanism with representation fine-tuning. GRUN significantly improves both unlearning effectiveness and model utility, supports sequential unlearning, and reduces training cost.Third, we identify a previously overlooked vulnerability in fine-tuning-based unlearning. We demonstrate that malicious users can craft stealthy forgetting requests that manipulate benign tokens into unlearning signals, thereby degrading model utility for normal users. To mitigate this risk, we introduce Scope-aware Unlearning (SU), which constrains the effective scope of forgetting signals by incorporating dynamically constructed scope samples into the training objective. SU substantially enhances robustness against such attacks while maintaining compatibility with existing unlearning frameworks.Collectively, this dissertation establishes a unified framework for data protection in generative AI. By connecting memorization analysis, unlearning methodology, and robustness evaluation, this work advances the development of generative models that are not only powerful but also trustworthy, responsible, and resilient to misuse.","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 (97 p.) : ill."]},{"key":"dc:title","label":"Title","values":["Data protection in generative AI : from data owners to model builders"]}]}],"canonical_facts":{"dc:contributor":["Tang, Jiliang","Xing, Yue","Liu, Hui","Tang, Xianfeng"],"dc:creator":["Ren, Jie"],"dc:date":["2026"],"dc:description":["Thesis (Ph.D.)--Michigan State University. Computer Science - Doctor of Philosophy, 2026","The rapid advancement of generative models, including text-to-image diffusion models and large language models (LLMs), has transformed content creation across modalities. However, these powerful models raise significant concerns regarding data protection, including copyright infringement, privacy leakage, unauthorized dataset usage, and the generation of harmful content. Addressing these risks requires systematic solutions that bridge the perspectives of both data owners and model builders. This dissertation presents a comprehensive study of data protection in generative AI, spanning memorization analysis, dataset auditing, unlearning enhancement, and robustness against malicious exploitation.First, we investigate memorization in text-to-image diffusion models through the lens of cross-attention mechanisms. We reveal that memorization is closely associated with disproportionate attention concentration on specific trigger tokens, leading to overfitting to training samples. Based on this insight, we propose efficient detection metrics grounded in attention entropy and layer-wise analysis, as well as inference-time and training-time mitigation strategies. Our methods significantly reduce memorization without degrading image quality or increasing computational overhead.Second, we address the challenge of removing copyrighted and privacy-sensitive data from LLMs through unlearning. We conduct a systematic analysis of fine-tuning-based unlearning methods and uncover a shared underlying mechanism between gradient-ascent-based and suppression-based approaches: both operate by distinguishing target data and suppressing related generations rather than truly erasing knowledge. Motivated by this finding, we propose Gated Representation UNlearning (GRUN), a general and parameter-efficient framework that combines a soft gating mechanism with representation fine-tuning. GRUN significantly improves both unlearning effectiveness and model utility, supports sequential unlearning, and reduces training cost.Third, we identify a previously overlooked vulnerability in fine-tuning-based unlearning. We demonstrate that malicious users can craft stealthy forgetting requests that manipulate benign tokens into unlearning signals, thereby degrading model utility for normal users. To mitigate this risk, we introduce Scope-aware Unlearning (SU), which constrains the effective scope of forgetting signals by incorporating dynamically constructed scope samples into the training objective. SU substantially enhances robustness against such attacks while maintaining compatibility with existing unlearning frameworks.Collectively, this dissertation establishes a unified framework for data protection in generative AI. By connecting memorization analysis, unlearning methodology, and robustness evaluation, this work advances the development of generative models that are not only powerful but also trustworthy, responsible, and resilient to misuse.","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 (97 p.) : ill."],"dc:identifier":["etd:53004","local:Ren_grad.msu_0128D_22007","https://doi.org/doi:10.25335/wv4s-kt58"],"dc:language":["eng"],"dc:relation":["Electronic Theses & Dissertations"],"dc:rights":["In Copyright"],"dc:subject":["Artificial intelligence"],"dc:title":["Data protection in generative AI : from data owners to model builders"],"dc:type":["text","Electronic dissertations"]},"updated_at":"2026-11-18T18:59:34Z"}