| 陈兵,成翔,张佳乐,谢袁源.联邦学习安全与隐私保护综述[J].南京航空航天大学学报,2020,52(5):675-684 |
| 联邦学习安全与隐私保护综述 |
| Survey of Security and Privacy in Federated Learning |
| 投稿时间：2020-05-30 修订日期：2020-08-10 |
| DOI：10.16356/j.1005-2615.2020.05.001 |
| 中文关键词: 计算机系统结构 联邦学习 模型安全 隐私保护 |
| 英文关键词:computer system structure federated learning model security privacy protection |
| 基金项目: |
|
| 摘要点击次数: 1250 |
| 全文下载次数: 2013 |
| 中文摘要: |
| 联邦学习是一种新型的分布式学习框架，它允许在多个参与者之间共享训练数据而不会泄露其数据隐私。但是这种新颖的学习机制仍然可能受到来自各种攻击者的前所未有的安全和隐私威胁。本文主要探讨联邦学习在安全和隐私方面面临的挑战。首先，本文介绍了联邦学习的基本概念和威胁模型，有助于理解其面临的攻击。其次，本文总结了由内部恶意实体发起的3种攻击类型，同时分析了联邦学习体系结构的安全漏洞和隐私漏洞。然后从差分隐私、同态密码系统和安全多方聚合等方面研究了目前最先进的防御方案。最后通过对这些解决方案的总结和比较，进一步讨论了该领域未来的发展方向。 |
| 英文摘要: |
| Federated learning is a novel distributed learning framework which enables the sharing of training data across multiple participants without compromising their data privacy. However， such novel learning mechanism can still suffer from unprecedented security and privacy threats from various attackers. This article mainly explores the security and privacy challenges of federated learning by first introducing the preliminary knowledge and threat models to facilitate understanding of the potential attacks. Second， three types of attacks launched by the internal malicious entities are summarized and meanwhile the security and privacy vulnerabilities of federated learning architecture are analyzed. Third， the state-of-art protection solutions in aspects of differential privacy， homomorphic cryptosystem， and secure multi-party aggregation are surveyed. Finally， by summarizing and comparing these solutions， the promising directions are discussed. |
| [HTML] 查看全文 查看/发表评论 下载PDF阅读器 |
| 关闭 |
