---
id: 20260307-T0-18
title: "FedEMA-Distill：联邦学习鲁棒蒸馏新方法"
title_en: "FedEMA-Distill: Robust Federated Learning Distillation"
url: https://ai.daily.yangsir.net/daily/20260307-T0-18
issue_date: 2026-03-07
publish_date: 2026-03-06T05:00:00.000Z
source_name: "arXiv cs.LG (ML)"
source_url: https://arxiv.org/abs/2603.04422
---

# FedEMA-Distill：联邦学习鲁棒蒸馏新方法

arXiv论文2603.04422提出指数移动平均引导的联邦学习蒸馏方法，解决客户端数据异构和恶意行为导致的性能下降问题。该方法通过动态调整模型更新权重，抑制客户端漂移，加速收敛速度。实验表明，在20%客户端恶意攻击的场景下，模型准确率仍保持87%，比传统联邦高15个百分点。该方法可直接应用于现有联邦框架，无需额外硬件支持。

## English Version

**FedEMA-Distill: Robust Federated Learning Distillation**

arXiv paper 2603.04422 proposes an EMA-guided federated learning distillation method to address performance degradation from client data heterogeneity and malicious behavior. By dynamically adjusting model update weights, it suppresses client drift and accelerates convergence. Experiments show 87% accuracy under 20% client attacks, 15 points higher than traditional federated learning. The method integrates directly into existing frameworks without additional hardware support.

---

**来源**：[arXiv cs.LG (ML)](https://arxiv.org/abs/2603.04422)

**详情页**：https://ai.daily.yangsir.net/daily/20260307-T0-18

---

*智语观潮 · Daily — https://ai.daily.yangsir.net/llms.txt*