Method of analyzing computer traffic based on recurrent neural networks

Abstract

The given paper proposes a method of analyzing network traffic based on recurrent neural networks. There overview of perspective approaches for analyzing network traffic in order to detect attacks is provided. The authors investigated the largest and currently the most relevant CICIDS2018 dataset. The methods of dealing with the class imbalance in a dataset by adapting the Focal Loss function to the problem of traffic analysis are considered. There proposed method provides the effective representation of information characteristics of network packets by means of encoder subnetworks. The resulting embeddings are fed at the input of the recurrent LSTM layer. The designed network meta-architecture is potentially effective for the presented dataset as well as for relevant analogues.