FCAR-Net:一种用于小样本网络流量的分类方法

Network traffic classification is very important for both network management and network security. In some cases, such as zero-day attacks, only limited network traffic samples can be obtained. Applying traditional network traffic classification methods in the above scenarios will cause serious overfitting problems, so few-shot traffic classification is required. Based on the idea of "learning to learn" in meta-learning, this paper proposes a traffic multi-classification model suitable for few shot question, called FCAR-Net. FCAR-Net is an end-to-end network traffic classification model. It learns features from raw traffic samples, performs model training through a meta-learning training strategy, uses a cross-attention module for feature weighting, and finally outputs classification results through a relational network. The model proposed in this paper introduces a cross-attention module to extract the correlations between samples that are ignored in traditional work. In addition, the paper details the preprocessing of the data. Finally, the effectiveness of FCAR-Net is verified in the traffic dataset. The experimental results show that the FCAR-Net model proposed in this paper shows the best results in all test tasks and has very good performance. In addition, through experiments It is verified that the FCAR-Net model can generalize well to unseen datasets and has good generalization ability.