Network Intrusion Detection Method Based on Attention Mechanism and DenseNet

Abstract

Convolutional neural network is very effective for feature extraction of network intrusion detection data. However, most of the existing intrusion detection methods based on convolutional neural networks are not deep in layers and when the neural network is deepened, problems such as vanishing gradients will occur. Aiming at the above problems, a network intrusion detection method that combines attention mechanism and DenseNet is proposed. The method turns pre-processed network traffic data into grayscale maps and extracts features using DenseNet, which can deepen the network while preventing gradient disappearance. In order to reduce the interference of invalid information in network traffic data and make the model extract important features better, the ECANet efficient channel attention mechanism is introduced to increase the weight of important features; and then the Swish activation function is introduced to further improve DenseNet, which makes the model have better robustness and accuracy. The method is able to perform deep feature learning, resist overfitting, and alleviate vanishing gradients. Experiments were conducted using the NSL-KDD dataset and the UNSW-NB15 dataset. The experiment shows that the model has improved in accuracy and F1-score metrics than other shallow models, which proves the effectiveness of the method.