Evaluating Automated Driving Planner Robustness against Adversarial Influence

TL;DR

This paper introduces a probabilistic trust model-based method to evaluate the robustness of autonomous driving planners against adversarial attacks, emphasizing the importance of targeted adversarial testing over generic safety assessments.

Contribution

It proposes a novel evaluation framework that estimates the difficulty for adversaries to induce unsafe behavior in machine learning-enabled planners, focusing on adversarial-specific protections.

Findings

Protection evaluations can identify vulnerabilities in camera-based object detectors.

Adversarial evaluation requires tailored approaches to effectively challenge protections.

The method helps quantify the robustness of autonomous vehicle planning systems against adversarial influence.

Abstract

Evaluating the robustness of automated driving planners is a critical and challenging task. Although methodologies to evaluate vehicles are well established, they do not yet account for a reality in which vehicles with autonomous components share the road with adversarial agents. Our approach, based on probabilistic trust models, aims to help researchers assess the robustness of protections for machine learning-enabled planners against adversarial influence. In contrast with established practices that evaluate safety using the same evaluation dataset for all vehicles, we argue that adversarial evaluation fundamentally requires a process that seeks to defeat a specific protection. Hence, we propose that evaluations be based on estimating the difficulty for an adversary to determine conditions that effectively induce unsafe behavior. This type of inference requires precise statements about threats, protections, and aspects of planning decisions to be guarded. We demonstrate our approach by evaluating protections for planners relying on camera-based object detectors.