A Malicious Traffic Detection Method Based on Attention Mechanism

At present, the malicious traffic detection method based on deep learning has become the mainstream method in this field. However, detection models using convolutional neural network often directly input the feature map formed after convolution operation to the next layer, without considering the different importance distribution of the feature map in channel dimension and spatial dimension, which leads to the failure of the model to learn more discriminative spatial feature representation of traffic data. This paper proposes a malicious traffic detection method based on attention mechanism. After convolution operation, attention modules are introduced from channel dimension and spatial dimension to emphasize important features and suppress secondary features, so as to enhance the feature learning ability of the model. Moreover, we use a traffic encoding model that can obtain context-dependent traffic encoding vectors. Comparative experiments show that our method has achieved better performance.