Title: ATLAS: A Comprehensive Benchmark for Assessing Adversarial Robustness in LiDAR Perception

Abstract:

While autonomous driving systems are routinely assessed using pristine dataset samples, their real-world viability hinges on resilience against unusual, structured, and potentially adversarial sensor disruptions. This vulnerability is particularly pronounced in LiDAR technology, where malicious actors can physically alter the sensing environment to cause black-box perception failures, all without needing direct access to the underlying algorithms. Currently, existing LiDAR benchmarks offer minimal insight into these specific failure scenarios. Previous research in adversarial LiDAR has predominantly focused on attack hardware, geometric and algorithmic countermeasures, and older detector generations, leaving the robustness of contemporary perception systems largely uninvestigated.

To bridge this critical evaluation gap, we present ATLAS (Adversarial Temporal LiDAR Attack Suite), the inaugural large-scale benchmark grounded in physical reality for testing LiDAR perception models against black-box sensor attacks. ATLAS simulates the two main categories of attacks—point injection and point removal—using authentic driving sequences. Our assessment of a diverse range of state-of-the-art LiDAR perception models uncovers a notable asymmetry in robustness: while models that excel on standard benchmarks demonstrate greater resistance to point removal attacks, they exhibit heightened susceptibility to point injection attacks compared to less performant models. We attribute this weakness to the use of standard object database sampling during data augmentation, highlighting how conventional training methodologies can trigger architecture-independent robustness failures. Furthermore, we explore preliminary strategies for mitigating both types of attacks. To facilitate extensible and reproducible evaluations as adversarial techniques advance, we have open-sourced the ATLAS generation code, aiming to establish black-box sensor robustness as a fundamental consideration in the future development of LiDAR perception systems.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC