分布式多方协作深度学习的安全与隐私保护

联邦学习和面向多数据中心的协作深度学习将是未来深度学习在现实中真正落地应用的主流范式。这种分布式的多方协作深度学习模式因暴露了更多的攻击面而面临诸多新的安全挑战和隐私威胁。本项目针对分布式多方协作深度学习中隐私攻击、拜占庭攻击和后门攻击的新模式，研究模型性能优化和数据隐私保护双重约束下的攻击检测和安全防御问题。内容包括：分布式多方协作深度学习中隐私泄露的深层机理与隐私保护模型；拜占庭模式下参与者隐私性和模型正确性双重目标的一致逼近；分布式多方协作深度学习的后门攻击防御和后门模型净化。本项目交叉融合密码学、博弈论和对抗学习等多领域理论与方法，在协作训练的各个关键环节全方位发力，突破其安全防御和隐私保护的系列关键技术，构建安全的分布式多方协作深度学习模型，并研发原型系统，为应用部署提供技术支撑。研究成果将助力于打通数据孤岛，在隐私保护法规范围内充分利用分散的小规模数据发展基于AI的服务和产业。

The multi-party cooperative deep learning, such as federated learning and collaborative deep learning for multiple data centers, will be the mainstream paradigm of deep learning in the future practical applications. In the real application environment, this cooperative mode faces many new security challenges and privacy threats because it exposes more vulnerable surfaces. This project focuses on the privacy attacks, Byzantine attacks and backdoor attacks aiming at multi-party cooperative deep learning, and aims to design the schemes for detecting attacks and defensing the models with the double constraints of model performance optimization and data privacy protection. The specific contents include: studying the deep mechanism of privacy disclosure and designing multi-party cooperative deep learning models with privacy preserving, designing multi-party cooperative deep learning models with both privacy preserving and high performance in the Byzantine environment, designing the methods to purify the backdoor models after cooperatively deep learning among multiple parties. In this project, we will synthetically utilize the theories and methods from cryptography, game theory and adversary learning to break through the key technologies, construct a secure multi-party collaborative deep learning model, and then develop a prototype system to provide technical support for application deployment. The contributions of this project will help to solve the problem of isolated data island and make full use of scattered small-scale data to develop AI based services and industries within the scope of privacy protection regulations.