Scholarlink

Federated learning (FL) and split learning (SL) are two prominent distributed learning approaches that enable model training without raw data sharing. SL offers better model privacy than FL by splitting the model architecture between clients and the server, making it preferable for resource-constrained environments. However, SL is slower due to relay-based training across multiple clients. This paper introduces federated split learning, FSLearning, which combines the strengths of both FL and SL, eliminating their inherent drawbacks. FSLearning integrates tensor regression to reduce communication costs and improve training efficiency. Our analysis and empirical results show that FSLearning achieves similar test accuracy and communication efficiency as SL while significantly reducing computation time for multiple clients. Empirical results show that FSLearning reduces transmitted parameters by up to 50% using ResNet3D. By incorporating tensor regression layers (TRLs), FSLearning compresses activations, enabling efficient Homomorphic Encryption (HE) integration. Privacy evaluations confirm that DP achieves the lowest total variation distance (TVD), reducing membership inference risks.

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

FSLearning: An Efficient Federated Split Learning Framework for Privacy-Preserving Disease Prediction

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Bin Li,
  • Xiaoqian Jiang,
  • Yu-Chun Hsu,
  • Arif O. Harmanci,
  • Hongchang Gao,
  • Xinghua Shi

摘要

Federated learning (FL) and split learning (SL) are two prominent distributed learning approaches that enable model training without raw data sharing. SL offers better model privacy than FL by splitting the model architecture between clients and the server, making it preferable for resource-constrained environments. However, SL is slower due to relay-based training across multiple clients. This paper introduces federated split learning, FSLearning, which combines the strengths of both FL and SL, eliminating their inherent drawbacks. FSLearning integrates tensor regression to reduce communication costs and improve training efficiency. Our analysis and empirical results show that FSLearning achieves similar test accuracy and communication efficiency as SL while significantly reducing computation time for multiple clients. Empirical results show that FSLearning reduces transmitted parameters by up to 50% using ResNet3D. By incorporating tensor regression layers (TRLs), FSLearning compresses activations, enabling efficient Homomorphic Encryption (HE) integration. Privacy evaluations confirm that DP achieves the lowest total variation distance (TVD), reducing membership inference risks.