Deep reinforcement learning-based intrusion detection in IoT system: a review
Hongyu Zhang, Carsten R. Maple · IET conference proceedings. · 2023
Due to the heterogeneity and high connectivity of smart devices, Internet of Things (IoT) network systems are becoming more complex and dynamic, which expands its attack surface. Further, a new generation of cyber attacks (including zero-day attacks) presents a significant threat to the IoT environment, while Adversarial Machine Learning (AML) attacks attempt to subvert detection methods. To tackle these challenges, modern intrusion detection systems (IDS) need to be self-adjusting and be resilient to AML to cope with changing environments and novel attacks (e.g. zero-day attacks). Reinforcement learning (RL) algorithms can provide self-adaptation and decision-making capabilities for the agent in the environment. Deep Reinforcement Learning (DRL) combines Deep Learning (DL) and Reinforcement Learning by using deep neural networks to represent the agent's decision-making policy for RL processing. Thus, DRL-based IDS can be considered as a feasible scheme for the dynamic IoT environment. This research provides a review of the current state-of-the-art DRL-based IDS and a critical evaluation of the methods involved. Based on this review paper, IoT security researchers will be able to quickly and efficiently understand, compare and organise the scattered research findings in the domain.