【摘要】 近年来，网络流量分类的数据安全问题备受关注。联邦学习能够在保障数据隐私的基础上，实现数据共享。目前，联邦学习在流分类上面临客户端数据不平衡的挑战。针对此问题，本文提出了一种基于客户端选择的不平衡联邦网络流量分类方法。针对多样本分布场景，设计了循环传递前置标签集模型获取标签类别指标，结合衡量样本平衡程度的权值散度指标和针对少数目客户端引入的安德鲁·耶奥协议下的客户端样本数目指标，计算指标综合得分，实现客户端选择。实验结果表明，与代表性文献方法相比，本文方法在不增加不平衡网络流量分类时间的情况下，F1分数可提高0.5%～10.0%。更多还原

【Abstract】 In recent years, the data security issue of network traffic classification has attracted much attention. Federated learning can achieve data sharing on the basis of ensuring data privacy. Currently, federated learning faces the challenge of client data imbalance in flow classification. To address this problem, this paper proposes an imbalanced federated network traffic classification method based on client selection. For multi-sample distribution scenarios, a circular transmission pre-label set model is designed to obtain label category indicators, combined with the weight divergence indicator to measure the degree of sample balance and the number of client samples under the Andrew Yeo protocol introduced for a small number of clients, comprehensive indicator scores are calculated, and client selection is implemented. Experimental results show that compared with representative literature methods, this method can increase the F1 score by 0.5% to 10.0% without increasing the time of imbalanced network traffic classification.更多还原