Byzantine-Robust Learning over Distributed and Decentralized Networks
报告题目：Byzantine-Robust Learning over Distributed and Decentralized Networks
报告专家：凌青（中山大学）
报告时间：12月5日，19:00-20:00
报告地点：腾讯会议 ID： 471-699-535会议密码：610065
报告摘要：Byzantine attacks are deadly threats to distributed and decentralized machine learning systems. Byzantine agents can manipulate their messages sent to honest agents and lead the learned models to oscillate, diverge, or converge to abnormal ones. In the first part of this talk, we highlight the impact of data heterogeneity to Byzantine-robust distributed machine learning. Such data heterogeneity includes that within each honest agent (inner variation), and that across the honest agents (outer variation). We develop a variance reduction method to eliminate the impact of inner variation, and a resampling strategy to reduce the impact of outer variation. In the second part of this talk, we show that naively extending the existing Byzantine-robust distributed algorithms to the decentralized regime does not necessarily work. We provide guidelines for devising favorable Byzantine-robust decentralized algorithms, and give a specific example that meets the guidelines.
专家简介：凌青，2001年与2006年于中国科学技术大学自动化系分别获得学士与博士学位，2006年至2009年担任密歇根理工大学电子工程与计算机科学系博士后研究员。2009年至2017年任教于中国科学技术大学自动化系，其间曾担任宾夕法尼亚大学电子与系统工程系访问学者、微软亚洲研究院铸星计划访问学者。2017年起担任中山大学计算机学院教授、博士生导师。主要研究方向为分布式优化与机器学习，研究工作获得IEEE信号处理协会青年作者最佳论文奖。现担任IEEE Signal Processing Letters杂志资深编辑。
邀请人：宋恩彬
