【摘要】 以深度学习为代表的机器学习技术已经在入侵检测方面取得显著进展，但对抗样本的出现会使入侵检测模型产生错误的结果，从而躲过检测，导致系统遭受恶意攻击。基于决策攻击的方法会进行多次查询，导致攻击容易被发现，而且效率较低。不同于传统的攻击方式，文中探索了一种针对入侵检测的频域对抗攻击，对入侵检测数据集进行傅里叶变换，利用低通滤波器，保留样本中更多的低频信息，去掉部分高频信息，再利用反傅里叶变换把修改后的数据转换回时域，实现基于频域的对抗攻击，从而检测入侵检测系统的鲁棒性。比较各种不同方法下生成的对抗样本与原始数据集攻击准确率，表明频域对抗攻击算法的攻击效果明显优于之前的对抗样本方法。更多还原

【Abstract】 Machine learning technology represented by deep learning has made remarkable progress in intrusion detection, but the appearance of adversarial examples will make the intrusion detection model produce wrong results, thus avoiding detection, resulting in malicious attacks on the system. The method based on decision attack will conduct multiple queries, making the attack easy to find and inefficient. Different from traditional attack methods, we explore a frequency adversarial attack for intrusion detection. The Fourier transform is performed on the intrusion detection data set, and a low-pass filter is used to retain more low-frequency information in the sample and remove part of high-frequency information, and then use the inverse Fourier transform to convert the modified data back to the time domain to implement frequency-based adversarial attacks, thereby detecting the robustness of the intrusion detection system. Comparing the attack accuracy of the adversarial examples generated by different methods with the original data set, frequency adversarial attack is better than the previous adversarial examples method.更多还原