Network Malicious Traffic Identification Method Based On WGAN Category Balancing

Aiming at the problem of data imbalance when in using deep learning model for traffic recognition tasks, a method of using Wasserstein Generative Adversarial Network (WGAN) to generate minority samples based on the image of the original traffic data packets is proposed to achieve a small number of data categories expansion to solve the problem of data imbalance. Firstly, through data preprocessing, the original traffic PCAP data in the dataset is segmented, filled, and mapped into grayscale pictures according to the flow unit. Then, the balance of dataset is achieved by using traditional random over sampling, WGAN confrontation network generation technology, ordinary GAN generation technology and synthetic minority oversampling technology. Finally, public datasets USTC- TFC2016 and CICIDS2017 are adopted to classify the unbalanced dataset and the balanced dataset on classic deep model CNN, and three evaluation indicators of precision, recall, and f1 are applied to evaluate classification effect. Experimental results show that the dataset balanced by the WGAN model is better than the ordinary GAN generation method, traditional oversampling method and the synthesis of the minority class sampling technique method under the same classification model.