ESE Ph.D. Thesis Defense: “OOD-Resilient Safety Monitoring for Safety-Critical Cyber-Physical Systems”

Out-of-distribution (OOD) data presents a significant challenge for data-driven runtime monitors of safety-critical cyber-physical systems. In particular, such monitors are rarely resilient to OOD data: both theoretical guarantees and empirical performance are difficult to maintain when the input to the data-driven component lies outside of the training distribution. In this thesis, we aim to develop a data-driven safety monitor for cyber-physical systems that overcomes these challenges. First, we present a safety monitor that achieves this resiliency through adaptive conformal prediction and incremental learning. The former allows for theoretical guarantees even on OOD data, and the latter boosts empirical performance on OOD data. Next, we consider potential alternatives to the building blocks of our safety monitor. As an alternative method for achieving empirical resiliency, we explore a technique for distribution shift reversal. We also present a diffusion model for forecasting highly-structured temporal sequences, which can be employed as the data-driven component of our safety monitor in the appropriate settings. Finally, we consider the practical applications of our safety monitor. We investigate extensions to make this monitor more applicable to real-world learning-enabled cyber-physical systems with rapidly changing distributions, and we apply the technique to medical cyber-physical systems.

Zoom link: https://upenn.zoom.us/j/99669942744