Model-Aware Generation of Safety-Critical Scenes for Robust End-to-End Driving via Distilled Planner Surrogates

Summary

A failure-mode analysis on NeuroNCAP finds that 95.9% of DiffusionDrive's collisions are planning failures despite reliable detection, which the authors attribute to rare safety-critical interactions in real driving logs. They propose a model-aware framework that generates 50 safety-critical scenes per planner, each pairing an attack with a safe escape trajectory, via a BEV-level differentiable surrogate distilled from the planner's own planning head. Fine-tuning on these scenes roughly halves the closed-loop collision rate of SparseDrive (28.0% to 12.2%) and DiffusionDrive (37.5% to 19.2%), with gains concentrated on the planner each scene set was generated against.

Related items

• InfoEvaluating the robustness of explainable AI methods to adversarial attacks in intrusion and malware detectionSimilar attack · OpenAlex (peer-reviewed AI security)
• InfoRL-ATTACK: A Generalized Black-Box Adversarial Attack on Language Models via Reinforcement LearningSimilar attack · Arxiv (cs.CR + cs.CL + cs.LG)
• InfoEvaluating Parameter-Efficient Fine-Tuning Attacks Against Non-Transferable Learning DefensesSimilar attack · Elsevier Security Journals
• InfoEscalate or Close: How Fast-Tier Models Handle Held-Out Web Attack Families in Two-Tier TriageSimilar attack · Arxiv (cs.CR + cs.CL + cs.LG)
• InfoVIPA: View-Invariant Projector-Based Adversarial AttackSimilar attack · Arxiv (cs.RO + cs.CV security)