Traffic prediction in SDN for explainable QoS using deep learning approach

Abstract

The radical increase of multimedia applications such as voice over Internet protocol (VOIP), image processing, and video-based applications require better quality of service (QoS). Therefore, traffic Predicting and explaining the prediction models is essential. However, elephant flows from those applications still needs to be improved to satisfy Internet users. Elephant flows lead to network congestion, resulting in packet loss, delay and inadequate QoS delivery. Recently, deep learning models become a good alternative for real-time traffic management. This research aims to design a traffic predicting model that can identify elephant flows to prevent network congestion in advance. Thus, we are motivated to develop elephant flow prediction models and explain those models explicitly for network administrators’ use in the SDN network. H2O, Deep Autoencoder, and autoML predicting algorithms, including XGBoost, GBM and GDF, were employed to develop the proposed model. The performance of Elephant flow prediction models scored 99.97%, 99.99%, and 100% in validation accuracy of under construction error of 0.0003952, 0.001697, and 0.00000408 using XGBoost, GBM, and GDF algorithms respectively. The models were also explicitly explained using Explainable Artificial Intelligence. Accordingly, packet size and byte size attributes need much attention to detect elephant flows.