Deep-learning-based Intrusion Detection with Enhanced Preprocesses

Intrusion detection has become a crucial issue due to an increase in cyberattacks. In most studies on this topic, intrusion detection performance has been found to be strongly related to the feature extraction and selection preprocess. However, there has been less research on problems or solutions related to the attributes of unequal metrics. Recently, deep-learning-based schemes have shown strong performance in image classification tasks without feature preprocessing. Therefore, in this study, we discuss the conversion of packet data into images for use in deep learning schemes with effective data preprocesses used to process the attributes of unequal metrics. A standard deviation standardization process is proposed to process the attributes of unequal metrics, which is followed by a data quantization process. Then, zigzag coding and the inverse discrete cosine transform are employed to convert the data into attribute images, which are used as the inputs for a convolutional neural network model. Intrusion detection is then achieved using the trained model. The experimental results demonstrate that the proposed scheme has reliable and efficient intrusion detection capability with a recall rate exceeding 94%. Meanwhile, packet attributes represented by 16 × 16 images provide about the same intrusion detection performance as that for 32 × 32 images. In summary, computational complexity can be reduced and performance can be maintained when using small images.