Scholarlink

<p>Machine learning conventionally assumes that training and testing data are centrally stored and follow the same distribution. However, in practical scenarios, strict privacy regulations necessitate decentralized data storage across heterogeneous devices and organizations, leading to significant distribution shifts. To tackle these challenges, Federated Test-Time Adaptation (FedTTA) integrates the global collaboration of Federated Learning (FL) with the dynamic adaptability of Test-Time Adaptation (TTA), enabling privacy-aware collaboration that continuously adapts to unseen data distributions. Despite growing interest and notable progress, current research remains fragmented and lacks a coherent conceptual foundation and unified taxonomy to consolidate emerging advances. In this paper, we provide a comprehensive survey of FedTTA, formalizing its problem setting and establishing a unified taxonomy encompassing three paradigms: i) Federated Initialization and Test Fine-tuning, where the global model serves as a robust prior for local refinement; ii) Federated Shared Backbone and Test Personalized Adaptation, where feature extraction is decoupled from lightweight local adapters; and iii) Federated and Test-time Collaborative Optimization, where a closed feedback loop is established for the continuous co-evolution of global and local models. Furthermore, we summarize representative datasets and applications. Finally, we highlight open challenges and discuss promising directions for future research. By offering a unified perspective and structured insights, this survey aims to advance the theoretical and practical understanding of FedTTA, steering the development of the next generation of adaptive, privacy-aware federated systems.</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

Federated Learning Meets Test-Time Adaptation: Methods, Challenges, and Future Directions

  • 期刊论文

  • 发表日期：

    2026年7月23日

查看全文
  • Chen Zhang,
  • Ge Su,
  • Huaxiao Zhou,
  • Lu Hao,
  • Fenglin Zhu,
  • Jintai Chen,
  • Yuxiang Cai,
  • Jianwei Yin,
  • Hongxia Xu

摘要

Machine learning conventionally assumes that training and testing data are centrally stored and follow the same distribution. However, in practical scenarios, strict privacy regulations necessitate decentralized data storage across heterogeneous devices and organizations, leading to significant distribution shifts. To tackle these challenges, Federated Test-Time Adaptation (FedTTA) integrates the global collaboration of Federated Learning (FL) with the dynamic adaptability of Test-Time Adaptation (TTA), enabling privacy-aware collaboration that continuously adapts to unseen data distributions. Despite growing interest and notable progress, current research remains fragmented and lacks a coherent conceptual foundation and unified taxonomy to consolidate emerging advances. In this paper, we provide a comprehensive survey of FedTTA, formalizing its problem setting and establishing a unified taxonomy encompassing three paradigms: i) Federated Initialization and Test Fine-tuning, where the global model serves as a robust prior for local refinement; ii) Federated Shared Backbone and Test Personalized Adaptation, where feature extraction is decoupled from lightweight local adapters; and iii) Federated and Test-time Collaborative Optimization, where a closed feedback loop is established for the continuous co-evolution of global and local models. Furthermore, we summarize representative datasets and applications. Finally, we highlight open challenges and discuss promising directions for future research. By offering a unified perspective and structured insights, this survey aims to advance the theoretical and practical understanding of FedTTA, steering the development of the next generation of adaptive, privacy-aware federated systems.