Survey on Byzantine attacks and defenses in federated learning

Federated learning as an emerging distributed machine learning, can solve the problem of data islands. However, due to the large-scale, distributed nature and strong autonomy of local clients, federated learning is extremely vulnerable to Byzantine attacks and the attacks are not easy to detect, which seriously damages the integrity and availability of the model. First, taking Byzantine attacks as the research object, a detailed classification and analysis of the attack principles were conducted. Secondly, guided by the classic network security defense model, federated learning defense methods were classified and analyzed from the perspective of defense mechanisms. Finally, the key issues and research challenges that need to be solved in federated learning to resist Byzantine attacks were proposed, providing new references for future relevant researchers.