Network Anomaly Detection Method Based on Joint Optimization of GAN and Classifier in Few-shot Scenarios

Abstract

Research into network anomaly detection methods in few-shot scenarios is of great practical significance, since there are relatively rare network anomaly events compared with normal ones. Sample augmentation based on GAN (Generative Adversarial network) has few requirements for domain knowledge and data types, and also has strong universality. In order to address the lack of guidance in hyperparameters optimizing for sample augmentation based on GAN, the joint optimization of GAN and classifier is introduced to realize the network anomaly sample augmentation and anomaly detection. Firstly, the network anomaly sample augmentation model and network anomaly detection classifier based on GAN are designed. Subsequently, hyperparameters trained with the sample augmentation model are optimized and selected according to the performance of the classifier on the augmented sample set. Finally, experiments are conducted on two types of datasets, namely network performance parameter characteristics and network traffic data characteristics. In accordance with the experimental results, the proposed sample augmentation method based on the joint optimization could effectively improve the accuracy of network anomaly detection in the few-shot scenarios.