Scholarlink

In order to solve the problem that adversarial sample attacks in network intrusion detection systems lead to a significant decrease in detection performance, this study proposes a multi-level defense framework based on an improved auxiliary classifier generative adversarial network (AC-GAN) and dynamic adversarial training. This paper constructs an adversarial traffic generator, designs a conditional generator based on the AC-GAN architecture, develops a feature purification module, uses autoencoders and spectral normalization convolutional layers to construct a traffic reconstruction network, and introduces an attention-guided feature selection mechanism to strengthen key protocol fields. A dynamic adversarial training strategy is further designed to alternately inject strong adversarial samples generated by projected gradient descent and diverse samples generated by AC-GAN during the training of the BiLSTM (Bidirectional Long Short-Term Memory) detection model. In the training phase, mixed precision training and gradient clipping are used to accelerate convergence. The optimizer uses RAdam and deploys cosine annealing scheduling to prevent local optimality. The average detection rate of AC-GAN is improved to 92.44%, while the false alarm rate is reduced and the detection time is reduced. This paper provides new ideas and technical support for network security protection.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Adversarial Sample Defense Mechanism for Network Intrusion Detection System Based on Generative Adversarial Network

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Jianyu Liu

摘要

In order to solve the problem that adversarial sample attacks in network intrusion detection systems lead to a significant decrease in detection performance, this study proposes a multi-level defense framework based on an improved auxiliary classifier generative adversarial network (AC-GAN) and dynamic adversarial training. This paper constructs an adversarial traffic generator, designs a conditional generator based on the AC-GAN architecture, develops a feature purification module, uses autoencoders and spectral normalization convolutional layers to construct a traffic reconstruction network, and introduces an attention-guided feature selection mechanism to strengthen key protocol fields. A dynamic adversarial training strategy is further designed to alternately inject strong adversarial samples generated by projected gradient descent and diverse samples generated by AC-GAN during the training of the BiLSTM (Bidirectional Long Short-Term Memory) detection model. In the training phase, mixed precision training and gradient clipping are used to accelerate convergence. The optimizer uses RAdam and deploys cosine annealing scheduling to prevent local optimality. The average detection rate of AC-GAN is improved to 92.44%, while the false alarm rate is reduced and the detection time is reduced. This paper provides new ideas and technical support for network security protection.