Dependable Lab

Designing a run-time monitor for timely and accurate detection of safety violations in CPS is challenging as it requires the precise modeling and analysis of the physical process, the environment, the cyber components that affect the physical process, and their interactions in both temporal and spatial domains. This research aims to investigate the fundamental problem of run-time safety assurance in autonomous CPS and propose a formal framework for the specification and design of context-aware hazard detection and mitigation mechanisms. An optimization approach is also proposed for further refinement of context-specific safety properties to capture the inter-scenario variability (e.g., different patient profiles or driving scenarios) and improve detection accuracy using a weakly supervised STL learning method.

Machine learning techniques are powerful to capture the input-output relations with high accuracy but should be constrainted to satisfy safety-related properties, especially in safety-critical applicatons (e.g., ADS, medical CPS). This work proposes a combined knowledge and data-driven approach for runtime prediction and mitigation of hazards. The proposed approach combines expert Knowledge on domain-specific Safety constraints with data from the closed-loop CPS operation to design a safety engine that can be integrated with a CPS controller’s interface to infer system context, predict impending hazards, and prevent the execution of unsafe control actions through generating preemptive and corrective actions.

In this work, we use an orthogonal model-driven approach to the regular data-driven techniques. Instead of focusing on exploring the entirety of the fault parameter space, we focus on a systematic characterization of the effect of the values of the parameter space (e.g., start time and duration of faults) in conjunction with the dynamic state of the vehicle to identify the most opportune system contexts to launch the attacks. We propose a Context-Aware safety-critical attack that can find the most critical context during a driving scenario to activate attacks that strategically corrupt the ADAS outputs, with the goal of (1) maximizing the chance of hazards and (2) causing hazards as soon as possible, before being detected/mitigated by the human driver or the ADAS safety mechanisms.