A Hybrid Neural Network Architecture for Early Detection of DDOS attacks using Deep Learning Models

With the increasing technologies in the digital era, the task of identification of Distributed Denial of Service (DDoS) attacks is a significantly challenging problem due to the computational complexity that has to be addressed. In practice DDos attack can be detected by classification of normal and abnormal traffic flow in network. Therefore hybrid neural network architecture for early detection of DDos attacks using deep learning models is proposed in this paper. As the most of the network systems against the DDoS attacks relies on pre-defined features taken by the entire traffic flows, an efficient method of early anomaly traffic detection is presented in this work in order to filter the abnormal traffic flow by auto profiling the traffic patterns which check only few of first bytes in every flow. Then two deep learning models named as Gradient Boosting Decision Tree (GBDT) parallel ensemble learning method and lightweight usable deep learning model which makes use of Convolutional Neural Network (CNN) properties are used for classifying spatial and temporal features of traffic flows respectively. Then, the outputs of these two models are merged using add function to combine the spatial and temporal features making an hybrid model for detecting the final traffic flow as either malicious or benign. This hybrid ensemble learning model demonstrated improved accuracy against existing related detection approaches.