联合学习
计算机科学
编配
分布式计算
MNIST数据库
计算机网络
复制(统计)
传播模式
分布式学习
删除
可靠性(半导体)
电信网络
数据建模
服务器
无线自组网
信息网络
人工智能
分布(数学)
作者
P. L. Theodoropoulos,Christos Mavrokefalidis,Kostas Berberidis
出处
期刊:
日期:2025-09-08
卷期号:: 1060-1064
标识
DOI:10.23919/eusipco63237.2025.11226769
摘要
Decentralized Federated Learning (DFL) is a server-less federated learning framework that facilitates collaborative model training while ensuring data privacy, eliminating the need for orchestration from a centralized entity. In many real-world applications, unreliable communication networks result in uneven agent participation due to stochastic environmental conditions. In this study, we model communication links between nodes as erasure channels, where transmitted information may be received by neighboring nodes with a low probability. To address this challenge, we propose two risk-aware extensions of the traditional Combine-then-Adapt (CTA) technique, employing the Conditional Value-at-Risk (CVaR) over the distribution of nodes. Our experimental evaluations on MNIST and FashionMNIST datasets show that the proposed approaches achieve higher performance as compared with standard CTA, across various setups.
科研通智能强力驱动
Strongly Powered by AbleSci AI