Scholarlink

<p>Typical Federated Learning (FL) solutions often rely on homogeneous model assumptions (e.g., identical architectures or aligned objectives), which fail to address the challenges of model heterogeneity, divergent objectives, and collaboration granularity in real-world scenarios. It significantly hampers the widespread adoption of FL across various application domains. To address the above challenging problems, we propose a novel and unified cloud-edge collaborative training framework, namely Prometheus. We design Prometheus as bilateral beneficial federated network architecture to address the model heterogeneity challenge and the model evolution problem. Specifically, for fair interaction between bilateral models and prevent greed-induced deception by edge clients, we design the Cloud-Edge interactive Handshake Protocol (CEHP) to facilitate standardized feature interaction on both sides, and propose the Information Gain Value Evaluation Mechanism (IGVE) to assess the fairness of cloud-edge models, which safeguard the privacy security of the Prometheus architecture. Furthermore, we propose a uniform and general scheme, namely the Attention Feature Fusion Mechanism (AFF), and innovatively introduce the Target Inference Hypothesis, achieving collaboration among objective heterogeneous models. Theoretical analysis demonstrates that the cooperation budget allocation during the CEHP process satisfies differential privacy. Extensive experiments on MNIST, CIFAR-10, and CIFAR-100 show that Prometheus achieves an average accuracy improvement of 3.2% over FedAvg and 4.1% over FedProx in non-IID settings, while maintaining rigorous privacy guarantees.</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

Prometheus: an efficient federated collaborative learning framework for coevolution of edge-cloud heterogeneous models

  • 期刊论文

  • 发表日期：

    2025年5月21日

查看全文
  • Zhigang Zhou,
  • Longzheng Dai,
  • Zengliang Bai,
  • Yiyou Dong

摘要

Typical Federated Learning (FL) solutions often rely on homogeneous model assumptions (e.g., identical architectures or aligned objectives), which fail to address the challenges of model heterogeneity, divergent objectives, and collaboration granularity in real-world scenarios. It significantly hampers the widespread adoption of FL across various application domains. To address the above challenging problems, we propose a novel and unified cloud-edge collaborative training framework, namely Prometheus. We design Prometheus as bilateral beneficial federated network architecture to address the model heterogeneity challenge and the model evolution problem. Specifically, for fair interaction between bilateral models and prevent greed-induced deception by edge clients, we design the Cloud-Edge interactive Handshake Protocol (CEHP) to facilitate standardized feature interaction on both sides, and propose the Information Gain Value Evaluation Mechanism (IGVE) to assess the fairness of cloud-edge models, which safeguard the privacy security of the Prometheus architecture. Furthermore, we propose a uniform and general scheme, namely the Attention Feature Fusion Mechanism (AFF), and innovatively introduce the Target Inference Hypothesis, achieving collaboration among objective heterogeneous models. Theoretical analysis demonstrates that the cooperation budget allocation during the CEHP process satisfies differential privacy. Extensive experiments on MNIST, CIFAR-10, and CIFAR-100 show that Prometheus achieves an average accuracy improvement of 3.2% over FedAvg and 4.1% over FedProx in non-IID settings, while maintaining rigorous privacy guarantees.