基于梯度安全性的联邦学习隐私技术研究

面对深度学习中大数据建模与数据孤岛分布的矛盾，联邦学习作为新兴的人工智能基础技术，通过不共享数据而共享梯度，在保护数据隐私的前提下提升了模型性能。然而梯度安全性问题阻碍了联邦学习的大规模应用。本项目研究基于梯度安全性的联邦学习隐私技术（压缩梯度的隐私分析、加噪梯度的隐私分析、基于向量同态加密的梯度保护），解决深度梯度泄露的可解释性、噪声等级和模型性能的平衡性、多密钥转换的安全聚合等关键问题，从而构建安全的联邦学习系统。本课题的研究有利于打破数据壁垒，在保护数据隐私和安全的前提下形成数据与知识共享的共同体，最终把人工智能的红利落实到社会的各个角落。

Facing the contradiction between big data modeling and data island distribution in deep learning, federated learning, as an emerging basic artificial intelligence technology, can improve the model quality under the premise of protecting data privacy by sharing the gradient but not sharing data. However, gradient security issues have hindered the large-scale application of federal learning. This project studies the federal learning privacy technologies based on gradient security, including privacy analysis of compressed gradients, privacy analysis of noise-added gradients, and gradient protection based on vector homomorphic encryption. By solving the key issues, such as the interpretability of deep gradient leakage, the balance between model performance and privacy budget, and security aggregation based on multi-key conversion, we can identify potential privacy vulnerabilities and build a secure federal learning system. The research of this subject is helpful to break down data barriers, form a community of data and knowledge sharing under the premise of protecting data privacy and security, and finally implement the dividends brought by artificial intelligence to all corners of society.