PPSplitFed: Privacy-Preserving Splitfed Learning via Functional Encryption

Jing Ma

doi:10.1109/tetc.2026.3683420

PPSplitFed: Privacy-Preserving Splitfed Learning via Functional Encryption

2026-04-24

0

PRE

AI

DOI：10.1109/tetc.2026.3683420

Original

Original request for help

Save

Abstract

En 中文

Federated learning (FL) and split learning (SL) are two emerging distributed learning mechanisms that enable model training without sharing the local data. The integration of FL and SL, called splitfed learning (SFL), reduces the training burden on the client side and parallelizes the model training to achieve efficient distributed model optimization without sharing the local data and the whole global model. Currently, concerns about the privacy leakage risks of splitfed learning have garnered much attention, but effective solutions have not yet been fully explored. We propose a novel privacy-preserving splitfed learning framework, named PPSplitFed, that guarantees the confidentiality of the intermediate parameters to protect data and model privacy. By designing a secure model training algorithm based on single-input functional encryption and a secure model aggregation algorithm based on multi-input functional encryption, PPSplitFed achieves secure model optimization. We describe the implementations of PPSplitFed on Multi-layer Perceptron (MLP) and Convolutional Neural Network (CNN) models and compare the performance with centralized learning (CL), FL, SL, SFL, and the state-of-the-art privacy-preserving SFL scheme, as well as privacy-preserving model training scheme on multi-sourced datasets. The experimental results show that PPSplitFed is effective and efficient with reduced computation and communication costs while maintaining the model accuracy.

Keywords:

Federated learning

split learning

splitfed learning

functional encryption

privacy protection

PPSplitFed: Privacy-Preserving Splitfed Learning via Functional Encryption

Abstract

Journal

Organization

Cited Papers