Dilated convolution based botnet detection model

Abstract

To address the problems of botnet stealthiness and difficulty in detection, this paper proposes a botnet detection model based on dilated convolution. The model first uses dilated convolution to increase the perceptual field of information and extract features from it, and then uses reflection padding to expand the extracted spatial features with samples, then uses squeeze-and-excitation networks to assign different weights to feature channels, and then uses gate recurrent unit to extract the temporal relationships preserved between features, and finally implements botnet detection. The model is validated on the UNSW-NB15 and CIC-IDS-2017 datasets with 99.4% and 99.3% accuracy, respectively, which verifies the effectiveness of the model for botnet detection.