[Paper Review] AutoRepair: Automated Repair for AI-Enabled Cyber-Physical Systems under Safety-Critical Conditions

Deyun Lyu, Jiayang Song+5 more|arXiv (Cornell University)|Apr 12, 2023

Smart Grid Security and Resilience4 citations

TL;DR

This paper proposes AUTOREPAIR, an automated framework that diagnoses and repairs unsafe control behaviors in deep neural network (DNN) controllers within safety-critical cyber-physical systems (CPS). Using optimization-based correction search and robust retraining, AUTOREPAIR identifies minimal control signal adjustments to satisfy system-level safety requirements, successfully repairing critical safety issues across six industrial CPS instances with significant reliability improvements.

ABSTRACT

Cyber-Physical Systems (CPS) have been widely deployed in safety-critical domains such as transportation, power and energy. Recently, there comes an increasing demand in employing deep neural networks (DNNs) in CPS for more intelligent control and decision making in sophisticated industrial safety-critical conditions, giving birth to the class of DNN controllers. However, due to the inherent uncertainty and opaqueness of DNNs, concerns about the safety of DNN-enabled CPS are also surging. In this work, we propose an automated framework named AutoRepair that, given a safety requirement, identifies unsafe control behavior in a DNN controller and repairs them through an optimization-based method. Having an unsafe signal of system execution, AutoRepair iteratively explores the control decision space and searches for the optimal corrections for the DNN controller in order to satisfy the safety requirements. We conduct a comprehensive evaluation of AutoRepair on 6 instances of industry-level DNN-enabled CPS from different safety-critical domains. Evaluation results show that AutoRepair successfully repairs critical safety issues in the DNN controllers, and significantly improves the reliability of CPS.

Motivation & Objective

• To address the growing challenge of unsafe behavior in DNN-enabled cyber-physical systems (CPS) deployed in safety-critical domains such as transportation and energy.
• To bridge the gap between detecting unsafe system behavior and automatically repairing DNN controllers, which remains underexplored despite advances in falsification and testing.
• To develop a general-purpose, automated repair framework that operates at the system level, ensuring safety requirements are met through optimal control signal corrections.

Proposed method

• Formulates the repair problem as a multi-objective optimization task to find minimal control signal corrections that satisfy safety requirements expressed in Signal Temporal Logic (STL).
• Uses iterative exploration of the control decision space to search for optimal corrections given an unsafe system execution trace.
• Applies optimization solvers to compute the minimal adjustments to control signals that restore safety, ensuring the repaired system execution satisfies all specified STL constraints.
• Integrates the repaired control inputs and outputs into a training dataset for robust retraining of the DNN controller.
• Implements the framework in Simulink, enabling compatibility with industrial CPS modeling and simulation workflows.
• Employs a feedback loop where repaired traces are used to retrain the DNN, enhancing its reliability and safety under real-world conditions.

Experimental results

Research questions

• RQ1 Can an automated framework identify and correct unsafe control behaviors in DNN controllers within safety-critical CPS without manual intervention?
• RQ2 How effective is the optimization-based correction search in minimizing control signal changes while ensuring system-level safety?
• RQ3 To what extent can the repaired DNN controller maintain performance and reliability after retraining with corrected execution traces?
• RQ4 How well does AUTOREPAIR generalize across diverse industrial CPS with different dynamics and safety requirements?

Key findings

• AUTOREPAIR successfully repaired critical safety issues in DNN controllers across six real-world industrial CPS instances from diverse safety-critical domains.
• The framework significantly improved system reliability by correcting unsafe control decisions through minimal, optimal adjustments to control signals.
• The retrained DNN controllers demonstrated enhanced robustness and safety compliance after incorporating repaired execution traces into the training process.
• The optimization-based repair method effectively identified minimal corrections that satisfied complex STL safety requirements without compromising system performance.