机器学习中的隐私保护技术

机器学习作为人工智能的核心技术已在多个领域得到应用。它在为人们带来便利的同时也暴露了严重的隐私泄露风险，面临训练数据被过度收集和滥用、训练过程中梯度数据被监听并用于倒推训练数据、参与训练方被污染后探知诚实参与方训练数据、所发布的模型接口和参数被利用来推断训练集成员等威胁。而机器学习的大数据驱动、多源异构、训练算法复杂、模型准确度要求高等特性，给隐私保护带来了巨大的挑战，使得现有的隐私保护方法难以适用。为此本项目拟基于机器学习和联邦学习两种典型场景，围绕机器学习各阶段的隐私泄露风险，深入探索机器学习的隐私保护理论与安全协议，解决隐私安全可度量、数据和模型安全可用、训练过程安全可控的问题。具体研究内容包括数据收集与处理阶段的隐私保护技术、机器学习训练阶段的隐私保护技术、联邦学习训练阶段的隐私保护技术、模型发布阶段的隐私保护技术。旨在突破隐私泄露对机器学习应用的制约，推动人工智能技术进一步发展。

Machine learning is the key technology for artificial intelligence and has been wildly applied in multiple fields. However, it reveals severe privacy and security problems while facilitating people's daily life, such as problems of training data over-collection, reconstruction by gradients spying, colluded participants attacks, attacks by observing prediction interface and parameters, etc. The data-driven, multi-sources, uninterpretable properties of machine learning technologies bring severe challenges for designing privacy-preserving mechanisms for machine learning models. Existing works can hardly solve the challenges. This project is going to design privacy-preserving technologies for classic centralized machine learning and federated learning against privacy leakage risks. It aims to deeply explore privacy theory and secure protocols for machine learning to achieve measurable privacy, usable privacy and controlled privacy. The main methods include privacy-preserving data processing and collection, private training, private federated training, and private model publishing. We attempt to break through the restraint of private and secure machine learning models and boost the development of artificial intelligence.