[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"doc-detail-124048-en":3,"doc-seo-124048-105":30,"detail-sidebar-cat-0-en-105":91},{"code":4,"msg":5,"data":6},0,"success",{"doc_id":7,"user_id":8,"nickname":9,"user_avatar":10,"doc_module":4,"category_id":11,"category_name":12,"doc_title":13,"doc_description":14,"doc_content":15,"file_id":16,"file_url":17,"file_type":18,"file_size":19,"view_count":20,"is_deleted":4,"is_public":20,"is_downloadable":20,"audit_status":20,"page_count":21,"language":22,"language_code":23,"site_id":24,"html_lang":23,"table_of_contents":25,"faqs":26,"seo_title":27,"seo_description":14,"update_tm":28,"read_time":29},124048,8796095461610,"Oliver","https://ap-avatar.wpscdn.com/davatar_276721f389ce27ea32af1340a28f341c",8,"Research & Report","Towards Comprehensive Security Assessment in Machine Learning Pipelines","Machine learning security is examined across the full ML pipeline, covering training data, model parameters, and model outputs. The dissertation analyzes data poisoning backdoor attacks and shows they can be effectively mitigated using simple fine-tuning, with proofs of non-robustness against straightforward fine-tuning strategies. It then investigates model stealing attacks by introducing new stealing methods for contrastive learning models and adaptive defenses. Finally, it addresses fake image detection and attribution for text-to-image generation, proposing prompt-based detection approaches.","Saarland University  \nDepartment of Computer Science  \nTowards Comprehensive Security Assessment in Machine Learning Pipelines  \nDissertation  \nzur Erlangung des Grades  \ndes Doktors der Ingenieurwissenschaften  \nder Fakultät für Mathematik und Informatik  \nder Universität des Saarlandes  \nvon  \nZeyang Sha  \nSaarbrücken, 2024  \nTag des Kolloquiums: 26.11.2024  \nDekan: Prof. Dr. Roland Speicher  \nPrüfungsausschuss:  \nVorsitzender: Prof. Dr. Ingmar Weber  \nBerichterstattende: Dr. Yang Zhang  \nProf. Dr. Michael Backes Prof. Dr. Tianhao Wang  \nProf. Dr. Xinlei He  \nAkademischer Mitarbeiter: Dr. Zheng Li  \nZusammenfassung  \nMaschinelles Lernen (ML) ist zu einem wesentlichen Bestandteil verschiedener kritischer Anwendungen geworden. Da ML-Modelle zunehmend eingesetzt werden, sind sie auch einer steigenden Anzahl von Angriffen ausgesetzt, die auf verschiedene Phasen der ML-Pipeline abzielen. Diese Pipeline kann grob in drei Phasen unterteilt werden: fürdas Training verwendete Daten, Modellparameter und die Ergebnisse des trainierten Modells.  \nIn dieser Arbeit führen wir eine gründliche Bewertung der Sicherheit maschinellen Lernens durch und untersuchen sie anhand dieser drei Phasen. Wir befassen uns zunächst mit der Datensicherheit und konzentrieren uns dabei insbesondere auf Backdoor-Angriffe durch Datenvergiftung. Wir zeigen, dass solche Angriffe mit einfachen Feinabstimmungsmethoden effektiv abgeschwächt werden können, und beweisen, dass die getesteten Backdoor-Angriffe nicht robust gegenüber einfachen Feinabstimmungsansätzen sind. Alsnächstes untersuchen wir die Modellsicherheit, indem wir Modelldiebstahlangriffe untersuchen. Ziel dieser Angriffe ist es, die Funktionalität eines Zielmodells mit minimalen Kosten und Rechenressourcen zu reproduzieren. Wir führen neuartige Modelldiebstahltechniken ein, die speziell auf kontrastive Lernmodelle abzielen, und entwickeln adaptive Abwehrmaßnahmen, um diesen Bedrohungen entgegenzuwirken. Abschließend wenden wir uns den Ergebnissen des Modells zu. Hier befassen wir uns mit der Erkennungund Zuordnung gefälschter Bilder und schlagen innovative Erkennungsmethoden vor, die Eingabeaufforderungen nutzen, um die Leistung zu verbessern. Dieser vielschichtige Ansatz ermöglicht es uns, die Sicherheit maschinellen Lernens aus einer umfassenden Perspektive anzugehen und alle kritischen Phasen der ML-Pipeline abzudecken.  \nAbstract  \nMachine learning (ML) has become an essential component in various critical applications. As ML models are increasingly deployed, they also face a rising number of attacks targeting different stages of the ML pipeline. This pipeline can broadly be divided into three phases: data used for training, model parameters, and the outputs of the trained model.  \nIn this dissertation, we conduct a thorough evaluation of machine learning security, examining it through the lens of these three stages. We begin by addressing data security, focusing particularly on backdoor attacks through data poisoning. We demonstrate that such attacks can be effectively mitigated with simple fine-tuning methods, proving that the backdoor attacks tested lack robustness against straightforward fine-tuning approaches. Next, we explore model security by investigating model stealing attacks. These attacks aim to replicate the functionality of a target model with minimal costs and computational resources. We introduce novel model stealing techniques specifically targeting contrastive learning models and develop adaptive defenses to counteract these threats. Lastly, we turn our attention to the outputs of the model. Here, we delve into the detection and attribution of fake images, proposing innovative detection methods that utilize prompts to enhance performance. These multifaceted approaches allow us to tackle machine learning security from a comprehensive perspective, spanning all critical stages of the ML pipeline.  \nBackground of this Dissertation  \nThis dissertation is based on the papers mentioned in","cbCaimfTIfhH2HvU","https://ap.wps.com/l/cbCaimfTIfhH2HvU","pdf",11878452,1,142,"English","en",105,"# Zusammenfassung\n## Maschinelles Lernen und Bedrohungen in der ML-Pipeline\n## Datensicherheit\n## Modellsicherheit\n## Erkennung und Zuordnung gefälschter Bilder\n# Abstract\n## Dreistufiger Rahmen der ML-Pipeline\n## Beitrag und Ergebnisse\n# Background of this Dissertation\n## Beiträge zu den drei Kernideen\n# Further Contributions of the Author","[{\"question\":\"该论文如何划分机器学习安全评估的范围？\",\"answer\":\"论文将ML管道大体分为三个阶段：用于训练的数据、模型参数以及训练后模型的输出。后续的研究分别围绕这三部分展开。\"},{\"question\":\"论文在数据层面的核心威胁是什么，如何应对？\",\"answer\":\"数据层面的重点是通过数据投毒实现的后门攻击。论文表明，这类攻击可以通过简单的微调有效缓解，并证明所测试的后门攻击对直接的微调策略缺乏稳健性。\"},{\"question\":\"论文如何处理模型参数与输出层面的安全问题？\",\"answer\":\"在模型参数层面，论文研究模型窃取攻击，并提出面向对比学习模型的新窃取技术与自适应防御措施；在输出层面，论文研究由文本到图像生成模型产生的伪造图像的检测与归因，并提出利用提示来提升检测性能的方法。\"}]","Towards Comprehensive Security Assessment in Machine Learning Pipelines | PDF",1785820093,358,{"code":4,"msg":31,"data":32},"ok",{"site_id":24,"language":23,"slug":33,"title":13,"keywords":34,"description":14,"schema_data":35,"social_meta":86,"head_meta":88,"extra_data":90,"updated_unix":28},"towards-comprehensive-security-assessment-in-machine-learning-pipelines","",{"@graph":36,"@context":85},[37,54,68],{"@type":38,"itemListElement":39},"BreadcrumbList",[40,44,48,51],{"item":41,"name":42,"@type":43,"position":20},"https://docshare.wps.com","Home","ListItem",{"item":45,"name":46,"@type":43,"position":47},"https://docshare.wps.com/document/","Document",2,{"item":49,"name":12,"@type":43,"position":50},"https://docshare.wps.com/document/research-report/",3,{"item":52,"name":13,"@type":43,"position":53},"https://docshare.wps.com/document/towards-comprehensive-security-assessment-in-machine-learning-pipelines/124048/",4,{"url":52,"name":13,"@type":55,"author":56,"headline":13,"publisher":58,"fileFormat":61,"inLanguage":23,"description":14,"dateModified":62,"datePublished":62,"encodingFormat":61,"isAccessibleForFree":63,"interactionStatistic":64},"DigitalDocument",{"name":9,"@type":57},"Person",{"url":41,"name":59,"@type":60},"DocShare","Organization","application/pdf","2026-08-04",true,{"@type":65,"interactionType":66,"userInteractionCount":20},"InteractionCounter",{"@type":67},"ViewAction",{"@type":69,"mainEntity":70},"FAQPage",[71,77,81],{"name":72,"@type":73,"acceptedAnswer":74},"该论文如何划分机器学习安全评估的范围？","Question",{"text":75,"@type":76},"论文将ML管道大体分为三个阶段：用于训练的数据、模型参数以及训练后模型的输出。后续的研究分别围绕这三部分展开。","Answer",{"name":78,"@type":73,"acceptedAnswer":79},"论文在数据层面的核心威胁是什么，如何应对？",{"text":80,"@type":76},"数据层面的重点是通过数据投毒实现的后门攻击。论文表明，这类攻击可以通过简单的微调有效缓解，并证明所测试的后门攻击对直接的微调策略缺乏稳健性。",{"name":82,"@type":73,"acceptedAnswer":83},"论文如何处理模型参数与输出层面的安全问题？",{"text":84,"@type":76},"在模型参数层面，论文研究模型窃取攻击，并提出面向对比学习模型的新窃取技术与自适应防御措施；在输出层面，论文研究由文本到图像生成模型产生的伪造图像的检测与归因，并提出利用提示来提升检测性能的方法。","https://schema.org",{"og:url":52,"og:type":87,"og:title":13,"og:site_name":59,"og:description":14},"article",{"robots":89,"canonical":52},"index,follow",{"doc_id":7,"site_id":24},{"code":4,"msg":5,"data":92},[93,97,101,105,110,115,120,123,128,131,135],{"id":20,"doc_module":4,"doc_module_name":46,"category_name":94,"show_sort_weight":95,"slug":96},"Story & Novel",90,"story-novel",{"id":47,"doc_module":4,"doc_module_name":46,"category_name":98,"show_sort_weight":99,"slug":100},"Literature",80,"literature",{"id":53,"doc_module":4,"doc_module_name":46,"category_name":102,"show_sort_weight":103,"slug":104},"Exam",70,"exam",{"id":106,"doc_module":4,"doc_module_name":46,"category_name":107,"show_sort_weight":108,"slug":109},5,"Comic",60,"comic",{"id":111,"doc_module":4,"doc_module_name":46,"category_name":112,"show_sort_weight":113,"slug":114},6,"Technology",50,"technology",{"id":116,"doc_module":4,"doc_module_name":46,"category_name":117,"show_sort_weight":118,"slug":119},7,"Healthcare",40,"healthcare",{"id":11,"doc_module":4,"doc_module_name":46,"category_name":12,"show_sort_weight":121,"slug":122},30,"research-report",{"id":124,"doc_module":4,"doc_module_name":46,"category_name":125,"show_sort_weight":126,"slug":127},9,"Religion & Spirituality",20,"religion-spirituality",{"id":126,"doc_module":4,"doc_module_name":46,"category_name":129,"show_sort_weight":126,"slug":130},"World Cup","world-cup",{"id":132,"doc_module":4,"doc_module_name":46,"category_name":133,"show_sort_weight":132,"slug":134},10,"Lifestyle","lifestyle",{"id":136,"doc_module":4,"doc_module_name":46,"category_name":137,"show_sort_weight":106,"slug":138},19,"General","general"]