精读笔记 - Attack of the Tails: Yes, You Really Can Backdoor Federated Learning

精读笔记 - Attack of the Tails: Yes, You Really Can Backdoor Federated Learning

1. 基本信息

论文标题	Attack of the Tails: Yes, You Really Can Backdoor Federated Learning
论文作者	Hongyi Wang, Kartik Sreenivasan, Shashank Rajput, Harit Vishwakarmaw Saurabh Agarwal, Jy-yong Sohn, Kangwook Lee, Dimitris Papailiopoulos
科研机构	University of Wisconsin-Madison, Korea Advanced Institute of Science and Technology
录用会议	NeurIPS 2020
摘要概括	此项工作围绕着联邦学习后门鲁棒性问题提出 edge-case backdoor。所谓 edge-case 作为恶意客户端的毒样本出现在联邦学习系统中所有客户端训练集中是小概率事件。恶意客户端结合该毒样本与良性样本，使用 PGD 算法更新迭代本地模型参数得到恶意模型更新并上传给服务器，从而实现全局模型后门