Subnet Level Anomaly Detection at Time Units Using VAE

Authors:

Abstract
With the rise in cyberattacks in recent years, research on Intrusion Detection Systems (IDS) has become increasingly active. In this study, we propose a novel IDS that detects anomalies in network traffic and estimates the source subnet of an attack by applying a Variational Autoencoder (VAE). During the training phase, normal traffic data is segmented into time units, and feature vectors are generated for each subnet. These feature vectors are then trained using the VAE. Subsequently, feature vectors are generated from traffic data containing anomalies using the same method as in the training phase. The generated feature vectors are reconstructed using the trained VAE, and subnets with high reconstruction errors are identified as the source sub nets of the attack. This method performs training and reconstruction for each subnet at each time unit, enabling the detection of the source sub net of an attack with low processing cost. In our experiment, we trained on the Monday dataset of CICIDS2017 and detected attacks in the datasets from Tuesday to Friday. Additionally, we implemented the reconstruction process of this method on a Field Programmable Gate Array (FPGA) and evaluated its performance.