报告题目 Title: CRAFT: Conflict-Resolved Aggregation for Federated Training

报告人 Speaker: Ziqi Wang 王子琦

报告人所在单位 Affiliation: Friedrich-Alexander-Universität Erlangen-Nürnberg 埃尔朗根-纽伦堡大学

主持人 Host：王玥

时间 Time:  2026-07-20 10:00-10:45

地点 Venue: 腾讯会议Online，会议号：866-1381-9470，密码：200433

报告摘要 Abstract: Federated learning enables collaborative model training across distributed clients without sharing private data, but its performance is often limited by statistical and system heterogeneity. Conventional weighted averaging may produce a global update that improves the overall objective while conflicting with individual client directions, degrading local performance and increasing client-level disparities. In this talk, we present CRAFT, a conflict-resolved aggregation framework for federated learning. Instead of weighted averaging, CRAFT formulates server aggregation as a geometric correction problem that seeks a global update close to a reference direction while maintaining positive alignment with all participating client updates. The resulting formulation admits a closed-form solution, avoiding iterative optimization and keeping server computation lightweight. We also discuss its layer-wise extension, its common-descent interpretation, and experimental results on heterogeneous benchmarks. CRAFT consistently improves both average accuracy and client-level fairness, providing a practical and robust solution for federated learning.

个人简介 Bio: Ziqi Wang is a Ph.D. student at Friedrich-Alexander-Universität Erlangen-Nürnberg (FAU) under the supervision of Prof. Enrique Zuazua. His research focuses on federated learning from a multi-level perspective, with particular interests in fair and robust server-side aggregation, client-side drift control and local optimization, game-theoretic incentive mechanisms, and privacy risks in distributed training.

海报 Poster: 王子琦 学术海报.jpg