基于GAN和元学习的伪装流量生成模型

打开文本图片集

关键词：生成对抗网络；恶意流量；对抗样本；元学习；黑盒攻击

中图分类号：TP393.08 文献标志码：A 文章编号：1671-6841（2026）01-0035-08

DOI：10. 13705/j. issn.1671-6841. 2024118

Abstract： The deep learning-based malicious traffc detection model is susceptible to adversarial attacks. In order to uncover security vulnerabilities with such models and find ways to enhance the robustness，an adversarial sample generation model （ReN-GAN） was proposed. Based on the principles of generative adversarial networks （GANs），the model could automatically generate relevant disguised traffic based on traffic features andutilizethe transferability of adversarial samples to achieve black-box attacks.By introducing momentum iteration methods and adding constraints on perturbations，the generalization capability of disguised traffic adversarial samples while ensuring the functionality of the original traffic was enhanced.During training，the model was optimized by integrating meta-learning theory，enabling the target integrated model to capture the common decision boundaries of various models more effectively and enhancing the transferability of generated adversarial samples.Experimental results showed that the adversarial samples generated by the ReN-GAN model，while preserving the characteristics of the original traffic ，achieved an average evasion rate of 54.1% on black-box detection models， significantly reducing the generation time compared to other methods.Furthermore，when trained on classfiers based on DNN， the ReN-GAN model required only five iterations to generate disguised trafic with an evasion rate of 62% ，greatly reducing the interaction times.

Key words： generative adversarial network； malicious traffic； adversarial samples； meta-learning;black-box attack

0 引言

近年来，随着机器学习技术的不断进步和普及，越来越多的检测系统采用了机器学习算法来检测恶意流量，增强了检测的速度及准确度[1-2]。（剩余12973字）