A Rule-Driven Approach for Safety-Violation Search in Autonomous Driving Systems

Autonomous driving systems (ADS) and robotic vehicles (RVs) have made significant advancements, yet safety and security challenges remain. Current approaches predominantly focus on external attacks and software vulnerabilities, while often neglecting critical safety concerns associated with the Safety of the Intended Functionality (SOTIF), which addresses faulty implementations and user misuse. This paper introduces RDA, a rule-driven fuzzing framework designed to uncover unexpected behaviors in ADS through the analysis of safety rules. RDA extracts these rules from vehicle manuals and translates them into Linear Temporal Logic (LTL) formulas for validation. The fuzzing engine mutates inputs based on these rules, using distance metrics to evaluate compliance with safety protocols. We validate RDA on ArduPilot, a widely-used open-source platform, and successfully identify misbehaviors linked to 28 extracted rules. This work uncovers previously unknown safety violations, contributing to the safety and reliability of ADS.