[论文解读] Practical Solutions for Machine Learning Safety in Autonomous Vehicles

Sina Mohseni, Mandar Pitale+2 more|arXiv (Cornell University)|Dec 20, 2019

Adversarial Robustness in Machine Learning参考文献 49被引用 51

一句话总结

本文综述了用于自主车辆的实用 ML 安全技术，将它们映射到 ISO 26262/SOTIF 的差距，并讨论了错误检测器和鲁棒性策略以提升可靠性。

ABSTRACT

Autonomous vehicles rely on machine learning to solve challenging tasks in perception and motion planning. However, automotive software safety standards have not fully evolved to address the challenges of machine learning safety such as interpretability, verification, and performance limitations. In this paper, we review and organize practical machine learning safety techniques that can complement engineering safety for machine learning based software in autonomous vehicles. Our organization maps safety strategies to state-of-the-art machine learning techniques in order to enhance dependability and safety of machine learning algorithms. We also discuss security limitations and user experience aspects of machine learning components in autonomous vehicles.

研究动机与目标

• Identify gaps in ISO 26262 and ISO/PAS 21448 (SOTIF) when applied to ML in autonomous vehicles.
• Catalog practical ML safety techniques that complement traditional software safety.
• Map ML safety techniques to engineering safety strategies to improve dependability.
• Present concrete implementations and discuss open challenges and future directions.

提出的方法

• Survey ML safety research on error detectors and model robustness.
• Organize techniques to align with safety strategies (Safe Fail, Safety Margins).
• Describe three practical implementations for safety-critical autonomous vehicle apps.
• Discuss run-time monitoring, uncertainty estimation, and OOD/detection methods.
• Analyze how these techniques address design specification, transparency, testing, and performance/robustness gaps.

实验结果

研究问题

• RQ1 What are the key safety gaps in ISO 26262 and SOTIF for ML components in autonomous vehicles?
• RQ2 What practical ML safety techniques can complement existing engineering safety standards?
• RQ3 How can ML safety techniques be mapped to established safety strategies (e.g., Safe Fail, Safety Margins)?
• RQ4 What concrete implementations illustrate safe deployment of ML in safety-critical AV applications?
• RQ5 What are the open challenges and future directions for ML safety in autonomous driving?

主要发现

• Safety gaps exist between ML properties and traditional software safety standards, particularly in design specification, transparency, testing, and performance under open-world conditions.
• Uncertainty estimation and run-time error detectors can enable safe-fail behavior in autonomous systems.
• OOD/unknown sample detection and domain generalization techniques improve robustness to distribution shifts.
• Robustness to corruptions, perturbations, and adversarial inputs enhances safety margins of ML components.
• Practical implementations demonstrate how monitoring and rejection mechanisms can be integrated into AV pipelines.