arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 2008.08330cs.CRcs.LGcs.NI
收藏

联邦学习网络的智能安全增强

Toward Smart Security Enhancement of Federated Learning Networks

University of Electronic Science and Technology of China(电子科技大学);PHENIKAA University(PHENIKAA大学);Nanyang Technological University(南洋理工大学)

Junjie Tan, Ying-Chang Liang, Nguyen Cong Luong, Dusit Niyato

摘要:

由于传统的集中式学习网络(CLN)在隐私保护、通信开销和可扩展性方面面临越来越多的挑战,联邦学习网络(FLN)已被提出作为支持机器学习(ML)模型训练的有前途的替代范例。与CLN中的集中式数据存储和处理相反,FLN利用许多边缘设备(ED)来分布式地存储数据和执行训练。通过这种方式,FLN中的ED可以在本地保存训练数据,这保护了隐私并减少了通信开销。然而,由于FLN内的模型训练依赖于所有ED的贡献,如果一些ED上传不正确或伪造的训练结果,即,中毒攻击在本文中,我们回顾了FLN的脆弱性,特别是中毒攻击和主流对策的概述。然而,现时的对策只能提供被动的保障,并没有考虑到教育署署长的贡献所支付的培训费用,导致培训成本不必要地偏高。因此,我们提出了一个智能安全增强框架FLN。特别是,一个验证前聚集(验证)的程序被开发来识别和删除非良性的训练结果从ED。然后,应用深度强化学习(Deep Reinforcement Learning,DRL)学习ED的行为模式,主动选择训练效果好、训练费用低的ED。仿真结果表明,该框架可以有效地保护FLN。

英文摘要:

As traditional centralized learning networks (CLNs) are facing increasing challenges in terms of privacy preservation, communication overheads, and scalability, federated learning networks (FLNs) have been proposed as a promising alternative paradigm to support the training of machine learning (ML) models. In contrast to the centralized data storage and processing in CLNs, FLNs exploit a number of edge devices (EDs) to store data and perform training distributively. In this way, the EDs in FLNs can keep training data locally, which preserves privacy and reduces communication overheads. However, since the model training within FLNs relies on the contribution of all EDs, the training process can be disrupted if some of the EDs upload incorrect or falsified training results, i.e., poisoning attacks. In this paper, we review the vulnerabilities of FLNs, and particularly give an overview of poisoning attacks and mainstream countermeasures. Nevertheless, the existing countermeasures can only provide passive protection and fail to consider the training fees paid for the contributions of the EDs, resulting in a unnecessarily high training cost. Hence, we present a smart security enhancement framework for FLNs. In particular, a verify-before-aggregate (VBA) procedure is developed to identify and remove the non-benign training results from the EDs. Afterward, deep reinforcement learning (DRL) is applied to learn the behaving patterns of the EDs and to actively select the EDs that can provide benign training results and charge low training fees. Simulation results reveal that the proposed framework can protect FLNs effectively and efficiently.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑