The network traffic detection system based on machine learning is a hot research direction in the field of network security at this stage, but at the same time, the network traffic detection system has been greatly challenged. The generation of attack samples reduces the detection performance of the detection system for malicious traffic. This article uses the generative adversarial network to generate adversarial samples, by adding noise interference to the original malicious traffic, so that non-directional disturbances that do not affect the characteristics of the original traffic are added to the attack characteristics to realize the judgment of the jamming detection model, thereby avoiding the characteristics detection and reducing the accuracy of traffic detection by 83.4%. It provides a richer training sample for the intrusion detection model to improve its robustness.