Scholarlink

Federated learning (FL) on edge devices often faces critical challenges, such as slow convergence rates and suboptimal performance, largely attributed to limited local data availability and data heterogeneity across participating devices. Recent advancements in generative artificial intelligence (AI) have demonstrated exceptional proficiency in synthesizing realistic data, offering potential solutions to these issues. In this paper, we present an innovative approach that enables edge devices to enhance their datasets by leveraging generative AI models. Recognizing varying resource capacities among edge devices, we propose a flexible data synthesis strategy, allowing each device to customize the quantity and quality of generated data according to its specific resource conditions and operational requirements. Building on this strategy, we formulate an optimization problem to maximize training accuracy while meeting the local resource constraints and task completion timelines. Experimental results reveal that our model can effectively enhance global accuracy compared to conventional FL algorithms, with even more pronounced improvements observed under skewed data distributions.

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

Efficient Federated Learning with Cost-Adjustable Generative AI over Heterogeneous Edge Devices

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Hanwen Zhang,
  • Peichun Li,
  • Jiawen Kang,
  • Liping Qian,
  • Yuan Wu,
  • Dusit Niyato

摘要

Federated learning (FL) on edge devices often faces critical challenges, such as slow convergence rates and suboptimal performance, largely attributed to limited local data availability and data heterogeneity across participating devices. Recent advancements in generative artificial intelligence (AI) have demonstrated exceptional proficiency in synthesizing realistic data, offering potential solutions to these issues. In this paper, we present an innovative approach that enables edge devices to enhance their datasets by leveraging generative AI models. Recognizing varying resource capacities among edge devices, we propose a flexible data synthesis strategy, allowing each device to customize the quantity and quality of generated data according to its specific resource conditions and operational requirements. Building on this strategy, we formulate an optimization problem to maximize training accuracy while meeting the local resource constraints and task completion timelines. Experimental results reveal that our model can effectively enhance global accuracy compared to conventional FL algorithms, with even more pronounced improvements observed under skewed data distributions.