Overview
This work examines the structural vulnerabilities of point cloud-based 3D object detectors, specifically focusing on how LiDAR-based detectors rely on spatial evidence in complex scenes. The research introduces an explainability-guided adversarial analysis methodology to understand and exploit these reliance patterns. It proposes the Saliency-LiDAR (SALL) method for generating universal saliency maps and the Explainability-aware Frustum Attack (EFA) which selectively perturbs influential frustums.
Research Context
Prior studies on adversarial robustness in 3D object detection have primarily focused on isolated 3D object models. Recent LiDAR spoofing attacks have targeted more realistic driving scenes but have emphasized physical realizability rather than gaining insights into detector behavior or attack efficiency. This research aims to address these gaps by investigating the efficiency of failures induced by exploiting detector reliance patterns on spatial evidence in complex scenes.
Approach
The research employed an explainability-guided adversarial analysis methodology. Key components of this approach include:
- Saliency-LiDAR (SALL) Method: This method aggregates Integrated Gradient attributions across scenes. Its purpose is to produce universal saliency maps for LiDAR-based 3D object detectors.
- Explainability-aware Frustum Attack (EFA): Guided by the saliency maps generated by SALL, EFA was designed to selectively perturb only the most influential frustums. This contrasts with approaches that uniformly attack entire object regions.
The methodology aimed to understand if exploiting detector reliance patterns could induce failures more efficiently.
Findings
Experiments were conducted on the KITTI and nuScenes datasets, evaluating detectors such as PointPillars and SECOND. The findings indicate the following:
- The Explainability-aware Frustum Attack (EFA) reduced detection recall by more than 15 percentage points.
- EFA achieved this reduction while requiring 25-50% fewer perturbed frustums compared to a state-of-the-art non-saliency-aware baseline.
- These results suggest that modern 3D detectors concentrate discriminative evidence within a small subset of spatial regions.
- This concentration of discriminative evidence exposes a structural robustness vulnerability in current LiDAR perception systems.
Why This Matters
The findings disclose that modern 3D detectors rely on a limited set of spatial regions for discriminative evidence. This specific reliance creates a structural robustness vulnerability within current LiDAR perception systems, impacting their reliability in complex scenes.
Potential Applications
The research identifies that the code for the Saliency-LiDAR method has been released.