Reinforcement learning and game theory based cyber-physical security framework for the humans interacting over societal control systems

A lot of infrastructure upgrade and algorithms have been developed for the information technology driven smart grids over the past twenty years, especially with increasing interest in their system design and real-world implementation. Meanwhile, the study of detecting and preventing intruders in ubiquitous smart grids environment is spurred significantly by the possibility of access points on various communication equipment. As a result, there are no comprehensive security protocols in place preventing from a malicious attacker’s accessing to smart grids components, which would enable the interaction of attackers and system operators through the power grid control system. Recently, dynamics of time-extended interactions are believed to be predicted and solved by reinforcement learning technology. As a descriptive advantage of the approach compared with other methods, it provides the opportunities of simultaneously modeling several human continuous interactions features for decision-making process, rather than specifying an individual agent’s decision dynamics and requiring others to follow specific kinematic and dynamic limitations. In this way, a machine-mediated human-human interaction’s result is determined by how control and physical systems are designed. Technically, it is possible to design dedicated human-in-the-loop societal control systems that are attack-resistant by using simulations that predict such results with preventive assessment and acceptable accuracy. It is important to have a reliable model of both the control and physical systems, as well as of human decision-making, to make reliable assumptions. This study presents such a method to develop these tools, which includes a model that simulates the attacks of a cyber-physical intruder on the system and the operator’s defense, demonstrating the overall performance benefit of such framework designs.