Privacy protection in federated learning: a study on the combined strategy of local and global differential privacy

With the increasing awareness of data privacy protection and the growing stringency of data security regulations, federated learning (FL) as a distributed machine learning approach has garnered widespread attention. However, in practice, FL faces severe challenges in privacy protection. This paper proposes a method that combines local differential privacy (LDP) and global differential privacy (GDP) in an FL environment. The algorithm addresses data privacy issues in FL while minimizing the impact on model performance. Our approach was evaluated on the MNIST and CIFAR-10 datasets, achieving superior results with 97.30% accuracy on MNIST and 90.18% accuracy on CIFAR-10, outperforming existing methods such as DPSaab and MPC. Additionally, LG-DPPA demonstrated strong resilience to privacy attacks, with a 97.44% membership inference attack (MIA) resistance on MNIST and a low adversarial attack success rate of 12.39% on CIFAR-10. The model maintained low utility loss, at only 1.91 on MNIST and 2.07 on CIFAR-10, highlighting its ability to balance privacy, accuracy, and computational efficiency. Additionally, the paper explores the impact of different privacy budgets and communication rounds on model performance, as well as how to choose appropriate parameter configurations in practical applications. This method is significant for handling sensitive data in FL environments and can also promote the development and application of privacy protection technologies.