Abstract
The security of cyber–physical systems is crucial due to their critical applications. The increasing success of machine learning (ML) has raised growing concerns about its impact on the cybersecurity of cyber–physical systems. Although several studies have assessed the cybersecurity of cyber–physical systems, there remains a lack of systematic understanding of how ML techniques can contribute to the use of deception on these systems. In this study, we aim to systematize findings on the use of ML for sensor data deception in both attack and defense scenarios. We analyzed 13 offensive and 3 defensive approaches that leverage ML for sensor data deception targeting cyber–physical systems. We summarized the offensive and defensive sensor data deception implementations with impact on cyber–physical systems at the system level, and the mechanisms to defend offensive deception. Additionally, we provide key insights and outline challenges intended to guide future research on defending against ML-based cyber deception in cyber–physical systems.