Trustworthy cyber-physical systems via physics-aware and AI-powered security

The rapid growth of cyber-physical systems (CPS), e.g., robotic vehicle (RV), has attracted extensive interest with their potential in numerous public and civilian applications. The safe operation of cyber-physical systems relies on the secure operation of both cyber and physical components. The frequency and sophistication of past attacks highlight the need for better defense mechanisms. Generally, a cyber-physical system faces unique challenges including internal control logic attacks and external physical interface attacks. Unfortunately, existing security solutions fail to provide sufficient protection to a real-time CPS because of the complicated sensor/actuator modules, closed-source controller software and constrained computing resources. This dissertation focuses on physical-aware and AI-powered security solutions for cyber-physical systems and more specifically embedded micro-controller systems (MCS). This dissertation presents cyber-physical security analysis and solutions through considering the interdependencies between the cyber and physical worlds. We propose data-driven analytical models combining the fine-grained cyber-physical analysis with machine learning techniques to secure the CPS in practical deployment scenarios. First, we propose a data-driven online monitoring framework that models the program-level control logic to detect data-oriented attacks against RVs. We leverage the internal dataflow information of RV controller functions to build a neural network-based approximate model as a secure replica to monitor the system operational behaviors at runtime. Next, we present a new stealthy attack against control state estimation-based detection techniques in RVs. We utilize data-driven analysis to search for vulnerable intermediate variables in controller functions. We leverage these variable-level interfaces in our learning framework to craft stealthy attacks and disrupt the safe operations of RVs. The techniques above motivate us to study the root causes of these controller vulnerabilities. Hence, we propose a firmware patching framework for RVs focusing on control-semantic bugs. This approach decompiles binary instructions and recovers controller functions to apply model-specific patches dedicated to the specific physical model. Apart from these security solutions, we also provide a privacy-preserving machine learning framework to deploy large deep neural networks on resources-constrained embedded devices in edge-cloud collaborations. This dissertation provides a solid step towards a secure and trustworthy cyber-physical system.Ph.D.Includes bibliographical reference