AdvFlowGAN: High Quality Adversarial Network Flow Synthesis Based on GAN

The evaluation of the adversarial robustness of intrusion detection systems (IDS) urgently requires high-quality adversarial flow samples. However, existing Generative Adversarial Network (GAN)-based generation methods face key challenges: modifications to functional features without constraints can easily lead to the loss of maliciousness in malicious samples, and the instability of model training severely restricts the quality of generated samples. Therefore, this paper proposes AdvFlowGAN, a high-quality traffic generation method for adversarial networks. This method innovatively distinguishes between functional and non-functional features of the flow and guarantees that the malicious nature of the generated samples remains unchanged by strictly restricting the scope of modification of functional features; at the same time, it introduces the spectral normalization technology to optimize the training process, which significantly improves the stability of the model and the realism of the generated samples. Experiments on public datasets show that AdvFlowGAN outperforms the benchmark method in terms of detection rate and escape increase rate. The method provides the security community an efficient tool for adversarial traffic generation and defense enhancement.