Enhancing Security, Safety, and Reliability of Modern Cyber-physical Systems

Hybrid Event: 3725 BBB / Zoom

Abstract: Modern cyber-physical systems (CPS), like autonomous driving systems and industrial control systems, showcase groundbreaking advancements driven by AI but face critical security and safety challenges. These challenges are amplified by the integration of complex, interconnected components like software controllers, AI models, and network systems, requiring cross-domain approaches to security analysis. Additionally, realistic assessments must account for intricate physical constraints, including resource limitations, timing requirements, and the laws of physics.

This talk highlights my research on enhancing CPS security, safety, and reliability across software, AI, and network layers, given the challenge of enforcing intricate physical constraints. I will demonstrate how realistic attack vectors exploiting AI weaknesses or software bugs, while adhering to these constraints, can lead to severe real-world consequences. My methodology leverages both empirical analysis and formal methods to model these physical constraints, and integrates this modeling with advanced security techniques that include adversarial machine learning, program analysis, and network system design.

Specifically, I designed software frameworks for identifying safety violations in autonomous driving and industrial control systems. These efforts leverage techniques from program analysis and formal methods to reason about compliance with safety policies under real-world physical constraints. Secondly, I investigate the robustness of AI-driven trajectory prediction components in autonomous vehicles, focusing on realistic adversarial scenarios that can lead to safety-critical outcomes such as hard braking or collisions. Thirdly, my research addresses security and reliability challenges in collaborative perception, where vehicles share sensor data to achieve perception tasks jointly. The projects improve the system’s robustness against asynchronous sensor inputs and adversaries fabricating data to share, considering realistic system-induced latencies, hardware limitations, and restricted adversarial knowledge. The above research outcomes are validated by high-fidelity simulations or real-world experiments, systematically advancing the trustworthiness of AI-powered CPS.