Attack Detection Based on Machine Learning Techniques to Safe and Secure for CPS—A Review

Abstract

Technological progression in communication and computing domains has led to the advent of cyber-physical systems (CPS). As an emerging technological advancement, CPS security is considered one of the prominent research directions these days. CPS is featured by its potential to integrate the cyber and physical data of the real world. CPS deployment in major infrastructure has shown the ability to reshape the world. Although, harnessing this ability is confined by their decisive nature and deep-seated consequences of cyber-attacks on surroundings or environment, infrastructure, and humans. In CPS, the substantial cyber concerns surge from the procedure of information transmission from multiple sensors to diverse actuators via the wireless medium, thus augmenting the attack region. Conventionally, CPS safety has been inspected from the standpoint of impending intruders from acquiring access to crucial systems using crypto-graphic or access control schemes. Thus, most research studies have emphasized attack detection in CPS. Although, in a sphere of growing adversaries, safe-guarding CPS from diverse adversarial attacks is becoming extremely sophisticated. Therefore, the need emerges for constructing resilient CPS which can con-front disruptions and stay functional despite adversarial attacks. Among the predominant methods investigated for constructing robust CPS, machine learning (ML) techniques have displayed greater suitability. However, from the latest studies regarding adversarial ML, it is advisable that for protecting CPS, ML techniques should themselves be robust. Therefore, this paper is intended at surveying the ML techniques employed for securing CPS and for detecting several attacks on CPS. It discusses the various design challenges, security objectives, security measures, security and reliability requirements of CPS, attack detection frameworks, and performance measures employed in prior works. Furthermore, it concludes with several research gaps and future directions for improving ML techniques and developing secure CPS.