A Survey of Federated Learning: Review, Attacks, Defenses

Traditional machine learning first processes the dataset and then, through training an initial model, eventually obtains the desired model. With the increasing awareness of data privacy and legal restrictions, the phenomenon of “data silos” has emerged. To overcome this issue, federated learning methods have been introduced. Federated learning emphasizes privacy protection and decentralization by allowing multiple participants to collaboratively train a model without sharing the raw data. However, recent research has shown that the training process in federated learning is also susceptible to various degrees of attacks, leading to data leakage. Therefore, researchers in the field have proposed various defense mechanisms for federated learning. In this paper, we introduce different types of federated learning, analyze some existing attacks and defense mechanisms in federated learning, and finally, we discuss future research directions.