An Efficient and Privacy-Enhanced U-Shaped Split Federated Learning Method for Distributed Multi-Client Systems

Machine learning (ML) plays a critical role in artificial intelligence (AI), particularly in image recognition tasks. Traditional ML relies on centralized model training, which requires substantial data transmission and storage, raising significant privacy concerns and limiting efficiency in resource-constrained devices or edge environments. Federated learning (FL) and split learning (SL) have emerged as promising distributed solutions to address these challenges. However, FL imposes a heavy computational burden on resource-limited devices, while SL suffers from slower training speeds due to its sequential nature. To address these limitations, we propose a distributed U-Shaped Split Federated Learning (USFL) method that allows a single server to manage multiple clients simultaneously. This approach combines the strengths of U-shaped SL and FL while incorporating a redesigned architecture. To further enhance privacy protection, USFL incorporates differential privacy mechanisms and noise injection strategies, ensuring robust data security. We evaluated USFL using the HAM10000 dataset and the ResNet-18 model to identify the optimal client-server configuration. Experimental results demonstrate that USFL outperforms centralized, SL, FL, and SFL methods in terms of model accuracy and runtime, making it a highly efficient and privacy-preserving solution for distributed learning tasks.