With the advent of the Internet of Things (IoT), network attacks have become more diverse and intelligent. In order to ensure the security of the network, Intrusion Detection system (IDS) has become very important. However, when met with the adversarial examples, IDS has itself become no longer secure, and the attackers can increase the success rate of attacks by misleading IDS. Therefore, it is necessary to improve the robustness of the IDS. In this paper, we employ Fast Gradient Sign Method (FGSM) to generate adversarial examples to test the robustness of three intrusion detection models based on convolutional neural network (CNN), long short-term memory (LSTM), and gated recurrent unit (GRU). We employ three training methods: the first is to train the models with normal examples, the second is to train the models directly with adversarial examples, and the last is to pretrain the models with normal examples, and then employ adversarial examples to train the models. We evaluate the performance of the three models under different training methods, and find that under normal training method, CNN is the most robust model to adversarial examples. After adversarial training, the robustness of GRU and LSTM to adversarial examples has greatly been improved.

随着物联网（IoT）的出现，网络攻击变得更加多样化和智能化。为了保证网络的安全，入侵检测系统（IDS）变得非常重要。然而，当遇到对抗样本时，IDS本身就变得不再安全，攻击者可以通过误导IDS来提高攻击的成功率。因此，有必要提高IDS的鲁棒性。在本文中，我们采用快速梯度符号方法 (FGSM) 生成对抗性示例来测试基于卷积神经网络 (CNN)、长短期记忆 (LSTM) 和门控循环单元 (GRU) 的三种入侵检测模型的鲁棒性。 ）。我们采用三种训练方法：第一种是用正常例子训练模型，二是直接用对抗样本训练模型，最后是用正常样本预训练模型，然后使用对抗样本训练模型。我们评估了三种模型在不同训练方法下的性能，发现在正常训练方法下，CNN 是对抗样本最稳健的模型。经过对抗性训练后，GRU 和 LSTM 对对抗性示例的鲁棒性有了很大的提高。