科学研究
学术报告
当前位置: 主页 > 科学研究 > 学术报告 > 正文

On the Robust Heterogeneous Federated Learning Algorithms

发布时间:2025-07-10 作者: 浏览次数:
Speaker: 陈川 DateTime: 2025年7月12日(周六)上午9:30-11:00
Brief Introduction to Speaker:

陈川现任中山大学计算机学院副教授、博士生导师。主要研究方向为可信联邦学习、鲁棒机器学习。在包括ICML, NeurIPS, ICLR, AAAI, IEEE TPDS, TKDE, TIFS等期刊会议发表论文80余篇,申请及授权专利20余项。主持国家重点研发计划项目子课题,NSFC面上项目、青年基金,CCF-腾讯犀牛鸟科研基金,腾讯微信犀牛鸟专项基金等多项课题,与微信/微众银行/网易游戏/招联金融/传音等多企业开展校企联合研究项目,部分成果完成技术转化并完成场景应用。获广东省计算机学会科技进步二等奖,广东省高校青年教师教学大赛二等奖,广东省高校教师教学创新大赛二等奖,CCF-腾讯犀牛鸟科研基金优秀奖,蚂蚁隐语优秀产学合作奖,IEEE Computer Society 2022 Best Paper Award (1 of 18 globally), World's Top 2% Scientists。

Place: 国交2号楼315报告厅
Abstract:Recent advances in federated learning (FL) have enabled collaborative model training across distributed, heterogeneous clients, driving innovations in privacy-sensitive domains like healthcare and finance. However, heterogeneity in data, models, and tasks introduces critical challenges, including biased aggregation, vulnerability to adversarial attacks, and communication inefficiency. This talk introduces some recent innovative methodologies to tackle these challenges, including: (1) tensor-driven optimization to capture complex cross-client correlations, (2) prototype alignment mechanisms to enforce domain-invariant feature representations.