Network Traffic Feature Engineering Based on Deep Learning

Aiming at extracting traffic features using manual selection and feature combination methods in current network traffic feature engineering, it is difficult to accurately extract the features of common traffic characteristics. A network traffic feature extraction method based on autoencoder model is proposed. The method first converts the first 144 bytes of the network data packet into a numeric code, and then acts as an input to the stacked autoencoder, then outputs a 49-dimensional feature through a 4-layer network encode. Using the dataset collected in laboratory to verify the method, experiments show that the feature extracted by this method can effectively extract network traffic characteristics, the extracted features are representative, and can use low-dimensional data to represent high-dimensional data.