SA-WGAN-Based Optimization Method for Network Traffic Feature Camouflage

Abstract

In the field of network security attack and defense, attackers frequently utilize network monitoring to analyze traffic features and obtain user privacy. Most defense methods employ feature-based traffic morphing techniques. However, the existing traffic camouflage method based on the Wasserstein Generative Adversarial Network (WGAN) exhibits limited defensive effectiveness, as the transformed traffic can still be detected due to the inherent limitations of the model and algorithm. In this paper, we propose a Wasserstein Generative Adversarial Network model with a Self-Attention mechanism (SA-WGAN) and adjust the parameter of the discriminator. Simultaneously, in the traffic generation algorithm, a constraint on padding packets was added: the Jaccard Index of the set of statistical features of the generated traffic must reach a threshold of 0.25, while ensuring the inclusion of three-way handshake packets to optimize the camouflage effect. To verify the camouflage effectiveness of the optimized defense method, we conduct a series of adversarial attacks. Experimental results show that the feature defense method based on SA-WGAN can significantly reduce the detection accuracy of monitored traffic. Compared to the feature defense method based on WGAN, it decreases the classification accuracy by 9.15% under the Panchenko attack, effectively enhancing the defensive capability and successfully increasing the difficulty for attackers to penetrate network traffic.