Collaborative Research: SaTC: CORE: Medium: Towards Secure Federated Learning

该项目将为新兴的联合学习范式提供安全基础。联合学习已经在不同的社会应用中大规模部署，因为它使许多客户端(例如智能手机、物联网设备和边缘设备)能够从机器学习模型中进行协作学习。在云服务器的帮助下，该过程可以在不必共享私人数据的情况下进行分析。虽然已经有很多关于提高联邦学习的准确性和通信效率的研究，但对其安全性的探讨却很少。在这个项目中，调查人员将通过探索对联合学习的新安全攻击以及开发新的安全联合学习方法来弥合这一差距，以降低分析和模型可能被外部参与者操纵的风险。该项目有三个目标，目标是联合学习的安全性。首先，研究团队将对联合学习的安全漏洞进行系统调查。特别是，他们将在联合学习的培训阶段探索安全漏洞，如中毒攻击和后门攻击。其次，该团队将开发可证明安全的联合学习方法，以防止中毒攻击和后门攻击。具体地说，将开发一些方法，确保有限数量的恶意客户端无论使用什么毒化和后门攻击，都不能以可证明安全的联合学习方法攻击机器学习模型。第三，研究团队将开发检测恶意客户端的方法，并有效地从攻击中恢复机器学习模型。调查人员将致力于现实世界的技术转移，将该项目的成果纳入新的和现有的本科和研究生课程，并开发和培训在开发安全的联邦学习系统方面具有重要经验的本科生和研究生研究人员，包括招收少数族裔和代表性不足的学生。该奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

This project will provide the security foundations for the emerging paradigm of federated learning. Federated learning has seen large-scale deployment in diverse societal applications because it enables many clients (e.g., smartphones, IoT devices, and edge devices) to collaboratively learn from a machine learning model. With help of a cloud server, the process allows analysis without having to share private data. While there are already many studies on improving the accuracy and communication efficiency of federated learning, its security is much less explored. In this project, the investigators will bridge the gap by exploring new security attacks to federated learning and developing new secure federated learning methods that reduce the risk that the analyses and models can be manipulated by outside actors. This project has three objectives targeting the security of federated learning. First, the research team will systematically investigate the security vulnerabilities of federated learning. In particular, they will explore security vulnerabilities in the training phase of federated learning, such as poisoning attacks and backdoor attacks. Second, the team will develop provably secure federated learning methods to prevent poisoning attacks and backdoor attacks. Specifically, methods will be developed that ensure a bounded number of malicious clients cannot attack the machine learning model in a provably secure federated learning method no matter what poisoning and backdoor attacks they use. Third, the team of researchers will develop methods to detect malicious clients and efficiently recover a machine learning model from attacks. The investigators will aim for real-world technology transfer, incorporate the results of this project in both new and existing undergraduate and graduate courses, and develop and train undergraduate and graduate researchers with significant experience for developing secure federated learning systems, including recruiting minority and under-represented students.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.