基于图神经网络的鲁棒加密流量识别

A robust encrypted traffic identification scheme based on graph neural network

当前的网络流量识别方法一般针对特定网络环境或数据集进行设计和测试,难以推广应用于复杂多变的实际网络环境。提出了一种基于图神经网络的鲁棒流量识别算法,用于在实际网络场景中实现准确的流量识别。首先,当前算法忽视网络环境波动,导致流行为模式发生改变,准确率下降,通过选择网络流中的高层协议特征对网络流进行聚类和筛选,以减小网络带宽波动对网站访问流量行为的影响。其次,由于当前算法大多只进行单流识别,忽视流间的相互关系,考虑网络流的多种类型特征信息及其相关性,并通过图神经网络提取网络流之间的时空相关特征,充分学习网络流量特征,通过多个流和多种特征的互补关系以提高算法的鲁棒性。最后,使用可以捕获数据全局信息的Transformer模型作为分类器对网络数据流的多类型特征进行分析,实现鲁棒网络流量识别。在不同网络环境下分别采集了对21个目标网站的共大约1500次和1400次的访问数据作为数据集进行训练测试,实现了90.7%的准确率,对比最新的ProGraph算法,准确率提高了7.3%,实验结果验证了所提方法的有效性。

The current methods for identifying network traffic are generally designed and tested for specific network environments or datasets,making it difficult to generalize and apply to complex and ever-changing actual network environments.A robust traffic recognition algorithm based on graph neural networks was proposed for achieving accurate traffic recognition in practical network scenarios.Firstly,in response to the current algorithm’s neglect of network environment fluctuations and the decrease in accuracy caused by pattern changes,network flows were clustered and filtered by selecting high-level protocol features to reduce the impact of network bandwidth fluctuations on website access traffic behavior.Secondly,due to the fact that most current algorithms only perform single stream recognition and ignore the interrelationships between flows,the various types of feature information and their correlations of network flows were considered,and spatiotemporal correlation features between network flows were extracted through graph neural networks to fully learn network traffic characteristics.By complementing multiple flows and features,the robustness of the algorithm was improved.Finally,a Transformer model that could capture global data information was used as a classifier to analyze the multi type features of network data flow,achieving robust network traffic recognition.Approximately 1500 and 1400 visits to 21 target websites in different network environments were collected as datasets for training and testing,achieving an accuracy of 90.7%.Compared with the latest ProGraph algorithm,the accuracy is improved by 7.3%,and the experimental results verify the effectiveness of the proposed method.