A Survey on Privacy Protection Techniques in Split Federated Learning

• Jie Wang,
• Qi Zhou,
• Xuechun Mao,
• Ying Chen,

Vol. 19, No. 10, October 31, 2025
10.3837/tiis.2025.10.017, Download Paper (Free):

• Federated learning
• Privacy Protection
• Split learning
• Split federated learning

Abstract

Split Federated Learning combines the decentralized training paradigm of Federated Learning with the model partitioning strategy of Split Learning, offering a promising framework for data privacy preservation. However, despite its advantages, recent research has demonstrated that this approach remains susceptible to privacy threats such as membership inference attacks, attribute inference attacks, and model inversion attacks. These vulnerabilities often stem from the leakage of intermediate features and gradients during collaborative training. This survey provides a comprehensive review of privacy-preserving methods within split federated learning, focusing on system architecture, typical attack pathways, and established defense mechanisms. In particular, it analyzes the implementation, effectiveness, and limitations of three major privacy protection techniques: differential privacy, homomorphic encryption, and secure multi-party computation. Furthermore, the paper discusses current challenges and future research directions, including dynamic model partitioning, hybrid defense strategies, and privacy preservation in non-independent and identically distributed data environments. This work aims to serve as a systematic reference and analytical framework for researchers exploring privacy protection in distributed machine learning.

Statistics

Cite this article