Develop a technique for generating reduced-order models of the protect surface of a cyber-physical-human critical infrastructure system enabling a comprehensive vulnerability assessment against catastrophic, model-based destabilization attacks.
Today's critical infrastructures are complex, dynamic, cyber-physical-human systems. These systems often can hide intricate sensitivities to small perturbations that can result in catastrophic, destabilizing behaviors, such as cascading failures in a power grid or a stock market's flash crash. Enemies with enough information about the system can exploit these sensitivities and design provably stealthy attacks to trigger them, so detecting the presence of these sensitivities, or these intrinsic vulnerabilities, is the first step towards protecting our critical infrastructure systems from this next-generation of sophisticated model-based attacks.
Many such model-based attacks in critical infrastructure systems are beginning to emerge . For example, machine learning can be used to design catastrophic attacks in a number of systems, such as in chemical processing plants, power generation or distribution systems, heating, ventilation, and air conditioning (HVAC) systems, water treatment or distribution systems, or nuclear power facilities. A general model may be known- from basic physics, but where specific parameters for a particular target facility are learned through stealthy observation of that system’s behavior. Nevertheless, even though details about designing and executing such attacks are increasingly available in the academic literature, little work has been done developing techniques to systematically detect and protect against them. Guaranteed robustness analyses  provide one approach to secure these systems, but the nonlinear and often hybrid nature of these systems, and their sheer complexity, make performing such computations extremely difficult at scale.
The protect surface of the critical infrastructure is a model that represents system variables that are hypothesized as being potentially exposed to possible attackers (or other unexpected perturbations), as well as their causal relationships to each other; dynamical structure functions have been used to build such models for linear time invariant systems. When building these models, choosing which variables are "exposed", and which variables are suppressed as part of the causal interaction between exposed variables, allows modelers to distinguish insider attacks, where many more system variables may be exposed, from other attacks where fewer variables may be exposed. Nevertheless, the number of exposed variables and the complexity of the, often nonlinear, dynamics can make these models unwieldy and impractical to develop for real critical infrastructure systems.
New research is needed to develop methodologies for reduced-order modeling of the protect surfaces for critical infrastructure systems, such as power systems, chemical and other manufacturing facilities, municipal and regional water systems, nuclear reactors, emergency services, transportation networks, pipelines, commercial and government facilities, financial systems, dams, communication networks, or food production and agricultural systems. These reduced order models should preserve critical properties of the full system, such as stability and sensitivity to perturbations of the exposed variables, while significantly reducing the complexity of the model. The reduced order model should exhibit the same vulnerability properties as the full model so that a comprehensive vulnerability analysis conducted on the reduced model will reveal the vulnerabilities and exploitation potential of the actual system. For more information on approaches for developing such reduced order models might be found in  and related works. A proposed solution should provide an approach to building the reduced models of a system, that can maintain the vulnerability properties of that system. It will provide an exemplar for the process, and a tool(s) to assist others in developing such models.