Scholarlink

<p>In real-world federated learning (FL) systems, dynamically joining clients often introduces unknown and heterogeneous data distributions, necessitating a rethinking of the initialization for these newly-joined client models and the communication efficiency across clients. Existing FL approaches typically initialize new client models by reusing the entire global model or a uniform subset of parameters, which constrains model adaptability and personalization under distributional shifts. In contrast, during the continuous evolution of biological populations, key species characteristics are encoded in their genes, enabling individuals to inherit and express superior genetic traits. Inspired by this mechanism, we propose <i>GENE-FL</i>, a gene-driven parameter-efficient dynamic federated learning framework that encapsulates generalized knowledge from diverse client groups into transferable and reusable neural components, termed <i>learnGene</i>s. The framework clusters clients by analyzing the principal components between data subspaces to identify distributional similarities. Within each cluster, task-sensitive layers are identified through a smooth local optimization process, enabling the generation of lightweight <i>learnGene</i>s for client–server interaction that substantially reduce communication overhead. For dynamically joining clients, the system assigns the most suitable <i>learnGene</i> based on their data distribution similarity to facilitate rapid adaptation to new data distributions. Extensive experimental results demonstrate that, compared to the classical FEDAVG, GENE-FL achieves a 4<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\times \)</EquationSource> </InlineEquation> reduction in communication cost, while requiring only about <i>9.04</i> MB of <i>learnGene</i> to effectively initialize dynamically joined client models.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Towards Adaptive and Communication-Efficient Dynamic Federated Learning

  • 期刊论文

  • 发表日期：

    2026年7月2日

查看全文
  • Shunxin Guo,
  • Jiaqi Lv,
  • Qiufeng Wang,
  • Xin Geng

摘要

In real-world federated learning (FL) systems, dynamically joining clients often introduces unknown and heterogeneous data distributions, necessitating a rethinking of the initialization for these newly-joined client models and the communication efficiency across clients. Existing FL approaches typically initialize new client models by reusing the entire global model or a uniform subset of parameters, which constrains model adaptability and personalization under distributional shifts. In contrast, during the continuous evolution of biological populations, key species characteristics are encoded in their genes, enabling individuals to inherit and express superior genetic traits. Inspired by this mechanism, we propose GENE-FL, a gene-driven parameter-efficient dynamic federated learning framework that encapsulates generalized knowledge from diverse client groups into transferable and reusable neural components, termed learnGenes. The framework clusters clients by analyzing the principal components between data subspaces to identify distributional similarities. Within each cluster, task-sensitive layers are identified through a smooth local optimization process, enabling the generation of lightweight learnGenes for client–server interaction that substantially reduce communication overhead. For dynamically joining clients, the system assigns the most suitable learnGene based on their data distribution similarity to facilitate rapid adaptation to new data distributions. Extensive experimental results demonstrate that, compared to the classical FEDAVG, GENE-FL achieves a 4 \(\times \) reduction in communication cost, while requiring only about 9.04 MB of learnGene to effectively initialize dynamically joined client models.