Detecting DDoS Attacks Using a New Polyscale Convolutional Neural Network for Policy Gradient Based DRL

Abstract

This paper presents a new architecture of a policy gradient based deep reinforcement learning (PGDRL) for detecting Internet traffic data (ITD) with distributed denial of service (DDoS) attacks (DDoS ITD) as unlabelled data. In this application, the main procedure in designing an intrusion detection system agent (IDSA) is policy approximation. Furthermore, a polyscale convolutional neural network (PCNN) is presented as a novel structure regarding the policy approximation. The IDSA aims to maximize its expected long-term rewards. Finally, the IDSA classification efficiency is assessed to find the detection rate of the proposed architecture. The PGDRL method detects the DDoS attack with almost 93% accuracy.