Collaborative Research: SaTC: CORE: Small: Securing Recommender Systems against Data Poisoning Attacks

协作研究：SaTC：核心：小型：保护推荐系统免受数据中毒攻击

• 批准号: 2125977
• 负责人: Neil Gong
• 金额: $ 40万
• 依托单位: Duke University
• 依托单位国家: 美国
• 项目类别: Standard Grant
• 财政年份: 2022
• 资助国家: 美国
• 起止时间: 2022-01-01 至 2024-12-31
• 项目状态: 已结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=2125977&HistoricalAwards=false
• 关键词: Collaborative; Research; SaTC; CORE; Small

The goal of this project is to build secure recommender systems against data poisoning attacks. Recommender systems are common online, suggesting movies, products, news, and many other kinds of items in order to help people find things they are interested in and make decisions. The influence recommender systems have on people's behavior, however, makes them attractive targets: attackers can create fake users who rate items in ways that lead the system to recommend products that are more in the attackers' interests than the users'. These "data poisoning" attacks threaten the integrity of recommender systems, harming both the companies and people that use them. This proposal will develop methods to detect, limit, and recover from the damage of data poisoning attacks, making recommender systems more resistant to manipulation by bad actors and thus more trustable and useful; the methods will also be incorporated into students' coursework and research work, training a next generation of computer scientists to build more robust machine learning systems.The project is structured around three main aims. Task 1 involves systematically investigating the security vulnerabilities of recommender systems against data poisoning attacks where attackers have varying levels of knowledge about the algorithms and datasets. In task 2 the team will develop new recommendation algorithms that provably prevent data poisoning attacks, i.e., a bounded number of fake users provably cannot affect the system's performance no matter how the fake users craft their rating scores. Task 3 is to develop methods to detect the fake users in data poisoning attacks with provable guarantees and efficiently recover a recommender system from data poisoning attacks. The project will provide research opportunities for students with backgrounds that are traditionally underrepresented in computing, and the work will be incorporated into courses at Duke University and West Virginia University and disseminated widely.This project is jointly funded by Secure and Trustworthy Computing (SaTC) and the Established Program to Stimulate Competitive Research (EPSCoR).This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.

这个项目的目标是建立一个安全的推荐系统来对抗数据中毒攻击。 推荐系统在网上很常见，推荐电影，产品，新闻和许多其他类型的项目，以帮助人们找到他们感兴趣的东西并做出决定。 然而，推荐系统对人们行为的影响使它们成为有吸引力的目标：攻击者可以创建假用户，这些假用户以导致系统推荐攻击者比用户更感兴趣的产品的方式对项目进行评级。 这些“数据中毒”攻击威胁到推荐系统的完整性，损害了使用它们的公司和用户。 该项目将开发检测、限制和恢复数据中毒攻击的方法，使推荐系统更能抵抗恶意行为者的操纵，从而更可靠和有用;该方法还将被纳入学生的课程和研究工作中，培养下一代计算机科学家，以构建更强大的机器学习系统。该项目围绕三个主要目标展开。 任务1涉及系统地调查针对数据中毒攻击的推荐系统的安全漏洞，其中攻击者对算法和数据集有不同程度的了解。在任务2中，团队将开发新的推荐算法，可证明可以防止数据中毒攻击，即，有界数量的假用户可以证明不会影响系统的性能，无论假用户如何制作他们的评级分数。任务3是开发方法来检测数据中毒攻击中的虚假用户，并提供可证明的保证，有效地从数据中毒攻击中恢复推荐系统。该项目将为具有传统上在计算领域代表性不足的背景的学生提供研究机会，这项工作将被纳入杜克大学和西弗吉尼亚大学的课程并广泛传播。该项目由安全和可信计算（SaTC）和刺激竞争研究的既定计划（EPSCoR）共同资助该奖项反映了NSF的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

项目成果

FedRecover: Recovering from Poisoning Attacks in Federated Learning using Historical Information

• DOI: 10.1109/sp46215.2023.10179336
• 发表时间: 2022-10
• 期刊: 2023 IEEE Symposium on Security and Privacy (SP)
• 作者: Xiaoyu Cao;Jinyuan Jia;Zaixi Zhang;N. Gong

Evading Watermark based Detection of AI-Generated Content

• DOI: 10.1145/3576915.3623189
• 发表时间: 2023-05
• 期刊: Proceedings of the 2023 ACM SIGSAC Conference on Computer and Communications Security
• 作者: Zhengyuan Jiang;Jinghuai Zhang;N. Gong

PoisonedEncoder: Poisoning the Unlabeled Pre-training Data in Contrastive Learning

• DOI: 10.48550/arxiv.2205.06401
• 发表时间: 2022-05
• 作者: Hongbin Liu;Jinyuan Jia;N. Gong

Almost Tight L0-norm Certified Robustness of Top-k Predictions against Adversarial Perturbations

• 发表时间: 2020-11
• 期刊: ArXiv
• 作者: Jinyuan Jia;Binghui Wang;Xiaoyu Cao;Hongbin Liu;N. Gong

PORE: Provably Robust Recommender Systems against Data Poisoning Attacks

• DOI: 10.48550/arxiv.2303.14601
• 发表时间: 2023-03
• 期刊: ArXiv
• 作者: Jinyuan Jia;Yupei Liu;Yuepeng Hu;N. Gong