Scholarlink

<p>Multimodal Federated Learning (MFL) leverages multimodal data from distributed clients to collaboratively train global models while preserving data privacy and exploiting the value of multimodal information. However, existing methods predominantly employ static fusion strategies and fixed hyperparameter configurations, which fail to address the high heterogeneity of clients in federated environments, thereby limiting model performance. To address this challenge, this paper proposes an Adaptive Multimodal Federated Learning framework via Multi-Armed Bandits for Fusion Strategy Selection and Hyperparameter Optimization (FedMAB). The framework unifies modal fusion strategy selection and hyperparameter optimization as online decision-making problems, achieving adaptive optimization through continuous learning of the mapping relationship between client characteristics and performance feedback. Specifically, the framework employs contextual multi-armed bandits to enable client feature-aware adaptive selection of fusion strategies, adopts a discretization-refinement mechanism to efficiently optimize critical hyperparameters including learning rate and contrastive learning temperature, and designs a cross-modal feature transformer to enable effective participation of unimodal clients in multimodal learning. Theoretical analysis demonstrates that the algorithm achieves sublinear regret bounds and convergence guarantees. Comprehensive experimental evaluation on datasets including Flickr30K and MS-COCO validates the superior performance and generalization capability of FedMAB compared to existing state-of-the-art methods. Ablation and convergence analysis experiments further confirm the effectiveness of each core component and the adaptive optimization capability of the framework.</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

FedMAB: adaptive multimodal federated learning with multi-armed bandits

  • 期刊论文

  • 发表日期：

    2026年1月3日

查看全文
  • Chaofeng Li,
  • Ruichun Gu,
  • Zhijian Dong

摘要

Multimodal Federated Learning (MFL) leverages multimodal data from distributed clients to collaboratively train global models while preserving data privacy and exploiting the value of multimodal information. However, existing methods predominantly employ static fusion strategies and fixed hyperparameter configurations, which fail to address the high heterogeneity of clients in federated environments, thereby limiting model performance. To address this challenge, this paper proposes an Adaptive Multimodal Federated Learning framework via Multi-Armed Bandits for Fusion Strategy Selection and Hyperparameter Optimization (FedMAB). The framework unifies modal fusion strategy selection and hyperparameter optimization as online decision-making problems, achieving adaptive optimization through continuous learning of the mapping relationship between client characteristics and performance feedback. Specifically, the framework employs contextual multi-armed bandits to enable client feature-aware adaptive selection of fusion strategies, adopts a discretization-refinement mechanism to efficiently optimize critical hyperparameters including learning rate and contrastive learning temperature, and designs a cross-modal feature transformer to enable effective participation of unimodal clients in multimodal learning. Theoretical analysis demonstrates that the algorithm achieves sublinear regret bounds and convergence guarantees. Comprehensive experimental evaluation on datasets including Flickr30K and MS-COCO validates the superior performance and generalization capability of FedMAB compared to existing state-of-the-art methods. Ablation and convergence analysis experiments further confirm the effectiveness of each core component and the adaptive optimization capability of the framework.