A survey of optimization algorithms for differential privacy in Federated Learning

Abstract

Federated Learning (FL), as a distributed machine learning approach, enables joint model training without sharing raw data, but during the model update process, the transmission of information still poses potential risks that may lead to the leakage of user privacy. In recent years, differential privacy (DP) techniques have been widely applied to federated learning to enhance data privacy protection. However, the introduction of differential privacy often has a negative impact on model performance, such as reducing model accuracy and increasing training time. Therefore, how to effectively balance privacy protection and model performance in federated learning has become an urgent problem to address. This paper first introduces the basic principles of federated learning and differential privacy, and then focuses on reviewing optimization algorithms for Differential Privacy in Federated Learning (DPFL). Unlike existing reviews on DPFL, we categorize the optimization algorithms into three types: noise mechanism optimization, privacy budget management, and model update optimization. By referencing a large number of related studies, we elaborate on the basic ideas, key innovations, and other aspects of various optimization methods, showcasing their performance and advantages in balancing privacy protection and model performance. Finally, we provide an outlook on future research directions, including further integrating DPFL with other advanced technologies to provide stronger support for applications in complex scenarios, enhancing visualization, and exploring the application of DPFL optimization algorithms in more practical fields.