https://blog.qife122.com/tags/%E8%81%9A%E5%90%88%E5%99%A8%E9%B2%81%E6%A3%92%E6%80%A7/ Recent content in 聚合器鲁棒性 on 办公AI智能小助手 Hugo zh-cn qife Fri, 12 Sep 2025 23:56:49 +0800 https://blog.qife122.com/p/%E5%88%86%E5%B8%83%E5%BC%8F%E5%BC%82%E6%9E%84%E6%95%B0%E6%8D%AE%E6%A0%87%E7%AD%BE%E4%B8%AD%E6%AF%92%E6%94%BB%E5%87%BB%E4%B8%8B%E5%9D%87%E5%80%BC%E8%81%9A%E5%90%88%E5%99%A8%E6%AF%94%E9%B2%81%E6%A3%92%E8%81%9A%E5%90%88%E5%99%A8%E6%9B%B4%E7%A8%B3%E5%81%A5/ Fri, 12 Sep 2025 23:56:49 +0800 https://blog.qife122.com/p/%E5%88%86%E5%B8%83%E5%BC%8F%E5%BC%82%E6%9E%84%E6%95%B0%E6%8D%AE%E6%A0%87%E7%AD%BE%E4%B8%AD%E6%AF%92%E6%94%BB%E5%87%BB%E4%B8%8B%E5%9D%87%E5%80%BC%E8%81%9A%E5%90%88%E5%99%A8%E6%AF%94%E9%B2%81%E6%A3%92%E8%81%9A%E5%90%88%E5%99%A8%E6%9B%B4%E7%A8%B3%E5%81%A5/ <h1 id="均值聚合器在分布式异构数据标签中毒攻击下比鲁棒聚合器更稳健">均值聚合器在分布式异构数据标签中毒攻击下比鲁棒聚合器更稳健</h1> <p><strong>摘要</strong><br> 针对恶意攻击的鲁棒性对分布式学习至关重要。现有研究通常考虑经典的拜占庭攻击模型，该模型假设某些工作节点可向服务器发送任意恶意消息，干扰分布式学习过程的聚合步骤。为防御此类最坏情况的拜占庭攻击，已提出多种鲁棒聚合器，并被证明有效且远优于常用的均值聚合器。</p>