An Effective Preprocess for Deep Learning Based Intrusion Detection

Abstract

The data preprocess directly affects the classification results in various applications. In the field of intrusion detection, less research raised the problems or solutions of unequal metrics in data attributes. This study proposes an effective data preprocessing method for network packets with unequal metrics in packet attributes. A standard deviation standardization was first applied to standardize each attribute of KDDCUP'99 dataset, followed by quantizing it to the range of 0 to 255 interval for afterward use of the image. Meanwhile, the Zigzag arrangement coding and IDCT (Inverse Discrete Cosine Transform) were then used to convert the quantized data into images. Experimental results demonstrate that a more than 94% recall rate of the overall intrusion detection classifier can be yielded by the proposed preprocess method even without a complicated network model. Meanwhile, intrusion detection performance can be guaranteed by using small-size images of packet attributes.