Privacy-preserved Federated Split Learning in Industrial Internet of Things
Zixuan Shu, Haitao Zhao, Bo Xu, Shen Qiao, Yukai Xu, Jinlong Sun, Qin Wang · 2024
With the development of the Industrial Internet of Things (IIoT), automated factory devices provide a large amount of data that can be used to train deep learning (DL) models. To protect the devices’ privacy while training the model, federated learning (FL) emerges as a promising framework, which enables devices to perform local training and transmit model parameters to the server. Nonetheless, FL confronts challenges due to devices’ limited computation and communication resources. Fortunately, split learning (SL) provides a feasible solution for FL through offloading specific computation tasks onto the server. This paper considers the federated split learning (FSL) framework, which integrates FL and SL. However, FSL continues to face the risk of privacy leakage. Therefore, an intermediate data perturbation mechanism based on differential privacy (DP) is introduced to address privacy concerns. To optimize the setting of the privacy budgets before the FSL training, we formulate a multi-objective problem considering the privacy protection and the performance of model training. Then, to seek the optimal solution to the problem, a generative adversarial network-powered genetic algorithm (GAN-GA) is proposed, which applies a generative adversarial network (GAN) in the genetic algorithm (GA). Experiments conducted on Cifar-10 demonstrate that GAN-GA successfully maintains a balance between model training and privacy protection.