A Systematic Literature Review on Malicious Use of Reinforcement Learning

Abstract

Since the inception of reinforcement learning (RL), there has been a growing interest in its application in various complex domains. Although these RL methods offer significant benefits of learning by their own experiences without an accurate system model, RL methods can also be used maliciously. This paper presents a systematic literature review of the state-of-the art RL-based cyberattacks to facilitate and motivate further research to address the potential RL misuse. We reviewed 30 recent primary papers and categorized them into (i) RL for attack planning, (ii) RL for performing intrusions, and (iii) RL for attack optimization. We also proposed an RL-based cyber attacks framework. Our insights on the status and limitations of the existing studies can help motivate related future studies.