Scholarlink

<p>The FL enables collaboratively training deep learning models on decentralized data. However, the current framework of FL, specifically the VFL framework, faces various challenges, such as subpar performance and safety concerns. Regrettably, these challenges have received limited attention within the existing research literature. In this study, we propose a novel VFL paradigm within VFL called AE-VFL, which improves the performance and safety of VFL by integrating gradient boosting into it. AE-VFL treats each local model within the participant as a sub-model and trains them sequentially to fit the gradient of the loss function derived from previous models. This approach leverages each sub-model to compensate for the limitations of its predecessors. Notably, each participant does not directly access the gradients of the ground-truth labels, which contain sensitive information. This mechanism ensures the performance and security of AE-VFL. Experimental results demonstrate that AE-VFL achieves superior performance and security compared to alternative approaches in a typical VFL setting. Furthermore, each participant benefits from AE-VFL’s utilization across various models and tasks. To facilitate the reproduction of this work, we have made the code for this project publicly available. You can access it via the following link: <a href="https://github.com/WOW5678/Towards-Efficient-and-Secure-Vertical-Federated-Learning-with-Additive-Ensemble">https://github.com/WOW5678/Towards-Efficient-and-Secure-Vertical-Federated-Learning-with-Additive-Ensemble</a>.</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 efficient and secure vertical federated learning with additive ensemble

  • 期刊论文

  • 发表日期：

    2026年2月16日

查看全文
  • Shanshan Wang,
  • Zhenxiang Chen,
  • Jianing Liu,
  • WenLin Xu

摘要

The FL enables collaboratively training deep learning models on decentralized data. However, the current framework of FL, specifically the VFL framework, faces various challenges, such as subpar performance and safety concerns. Regrettably, these challenges have received limited attention within the existing research literature. In this study, we propose a novel VFL paradigm within VFL called AE-VFL, which improves the performance and safety of VFL by integrating gradient boosting into it. AE-VFL treats each local model within the participant as a sub-model and trains them sequentially to fit the gradient of the loss function derived from previous models. This approach leverages each sub-model to compensate for the limitations of its predecessors. Notably, each participant does not directly access the gradients of the ground-truth labels, which contain sensitive information. This mechanism ensures the performance and security of AE-VFL. Experimental results demonstrate that AE-VFL achieves superior performance and security compared to alternative approaches in a typical VFL setting. Furthermore, each participant benefits from AE-VFL’s utilization across various models and tasks. To facilitate the reproduction of this work, we have made the code for this project publicly available. You can access it via the following link: https://github.com/WOW5678/Towards-Efficient-and-Secure-Vertical-Federated-Learning-with-Additive-Ensemble.