Deep Reinforcement Learning for intrusion detection in Internet of Things: Best practices, lessons learnt, and open challenges

The Internet of Things (IoT) scenario places important challenges even for deep learning-based intrusion detection systems. IoTs are highly heterogeneous networks in which multiple types of nodes and connections between them proliferate at a fast pace. From a deep learning perspective, such complexity translates into dynamic feature spaces where the extraction of semantic patterns and correlations among features may require sophisticated inductive biases to be learnt by gradient-based techniques. The research community has recently suggested using Deep Reinforcement Learning (DRL) as a potent approach to effectively identify cyber-threat attempts in IoTs.DRL consists of a Markov Decision Process-based meta-model that permits solving high-dimensional combinatorial optimization problems where differentiable supervisory signals may be absent. For this reason, multiple intelligent intrusion detection systems have been proposed for the IoT environment where high-level requirements are been pursued alongside the detection accuracy. These goals are related to optimizing the computational overhead, reducing power consumption at the edge, and preserving the privacy of sensitive information, among others.This survey offers a clear bird's eye view of the most recent design choices for DRL-based intrusion detection systems with a focus on the specific context of IoT. Our aim is not only to offer an exhaustive taxonomy of design alternatives made by DRL practitioners in the field of Intrusion detection, but also to discuss the advantages and the effective deployment of each setting concerning real IoT environments. We hope this work would guide the researchers interested in Intrusion Detection for IoTs to establish solid criteria for the most effective usage of Deep Reinforcement Learning in their future work.