Context-Aware Assurance in Cyber-Physical Systems

Rapid advances in sensing and computing technologies have led to the proliferation of Cyber-Physical Systems (CPS). However, the increasing use of connected and complex devices, shrinking technology sizes, and shorter time to market have increased the vulnerabilities of CPS to accidental and malicious faults, posing significant challenges in ensuring their reliability, safety, and security.
This dissertation presents a holistic approach to context-aware assurance in CPS through (i) control-theoretic specification of safety requirements and (ii) combined knowledge and data-driven refinement of safety specifications for run-time safety monitoring, hazard mitigation, and design-time safety validation.
As the foundation of this research, we propose a formal framework for the specification of safety context defined as the combinations of the cyber-physical system states, control actions, and potential hazards, based on a control-theoretic hazard analysis method. The safety context is specified using Signal Temporal Logic (STL) to consider the timing constraints for both hazard prediction and mitigation and consists of two parts: (i) the Unsafe Control Action Specification that describes the system states under which specific control actions are potentially unsafe and can eventually lead to hazards at a future time; and (ii) the Hazard Mitigation Specification that identifies the control actions that if issued by the controller within a specific time period can prevent potential hazards. An optimization approach is also proposed for further refinement of the context-specific safety properties to capture the inter-scenario variability (e.g., different patient profiles or driving scenarios) and improve detection accuracy. The final context-specific safety properties are then synthesized into the logic of a safety engine that can be integrated with a CPS controller as a wrapper with only access to the input and output data, and be used for run-time context inference, hazard prediction and mitigation, and safetyvalidation in different CPS that share the same functional specifications.
We propose combined knowledge and data-driven methods that integrate the generated safety context specifications or other safety constraints described as formal logic into machine-learning models for early hazard prediction and mitigation by enforcing the satisfaction of safety requirements while maintaining high prediction accuracy.
The generated safety context specifications can also be used for the safety validation of CPS at design time. We propose a model-driven approach orthogonal to the traditional data-driven techniques, which uses the system context specifications as the most opportune times for the activation of faults and efficiently identifies optimal fault values that can cause hazards as soon as possible without being detected by the existing safety mechanisms or mitigated by human interventions. The final goal of this approach is to discover potential design defects and safety-critical vulnerabilities in CPS to help with safety validation.
We evaluate the proposed approaches by developing open-source closed-loop testbeds that integrate real-world control software and physical-world simulators together with a fault injection engine that simulates the effect of accidental and malicious faults and real-world adverse events reported in the literature. We also evaluate our approaches using publicly available datasets or actual CPS. Experimental evaluation of the proposed assurance solutions for the case studies of artificial pancreas systems (APS) and advanced driver assistant systems (ADAS) demonstrates their generalization to a broad range of CPS with improved accuracy, timeliness, and robustness.