Abstract: Federated learning is a machine learning framework that enables participants in different fields to participate in largescale centralized model training together under the condition of protecting local data privacy. In the context of addressing the pressing issue of data silos, federated learning has rapidly emerged as a research hotspot. However, the heterogeneity of training data among different participants in federated learning also makes it more vulnerable to model robustness attacks from malicious participants, such as backdoor attacks. Backdoor attacks inject backdoors into the global model by submitting malicious model updates. These backdoors can only be triggered by carefully designed inputs and behave normally when input clean data samples, which poses a great threat to the robustness of the model. This paper presents a comprehensive review of the current backdoor attack methods and backdoor defense strategies in federated learning. Firstly, the concept of federated learning, the main types of backdoor attacks and backdoor defenses and their evaluation metrics were introduced. Then, the main backdoor attacks and defenses were analyzed and compared, and their advantages and disadvantages were pointed out. On this basis, we further discusses the challenges of backdoor attacks and backdoor defenses in federated learning, and prospects their research directions in the future.

Key words:

摘要： 联邦学习(federated learning, FL)是一种机器学习框架，能够使不同领域的参与者在保护本地数据隐私的条件下，共同参与大规模集中模型训练，在如今数据孤岛问题亟待解决的背景下迅速成为研究热点.然而，联邦学习中不同参与者之间训练数据具有异构性的特点，也使其更加容易受到来自恶意参与者的模型鲁棒性攻击，例如后门攻击.后门攻击通过提交恶意模型更新向全局模型注入后门，这些后门只能通过精心设计的输入触发，对模型鲁棒性造成极大的威胁.对联邦学习中目前的后门攻击方法及后门攻击的防御策略进行了全面综述.首先介绍了联邦学习的概念、后门攻击与防御的主要类型及其评价指标；然后分别对目前主要的后门攻击与防御方案进行了分析与比较，指出了它们的优势与不足；在此基础上，进一步讨论了联邦学习后门攻击与防御所面临的挑战，并展望了它们未来的研究方向.

关键词: