Research map / Detection and segmentation / 3D detection and driving
Towards Effective and Transferable Physical Camouflage against Multi-View BEV-based 3D Perception in Autonomous Driving
Linye Lyu, Jiawei Zhou, Daojing He, YU LI
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes a physical camouflage attack against multi-view BEV-based 3D perception models in autonomous driving, aiming to be effective and transferable across models.
Problem
Robustness of multi-view BEV-based 3D perception models in autonomous driving against physical adversarial camouflage attacks is largely unexplored; existing camouflage attacks target single-view 2D detectors.
Approach
Designs a physical camouflage attack targeting multi-view BEV 3D perception rather than single-view 2D detectors, with the goal of effectiveness and transferability across models. Full method details are unavailable from the abstract.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| apply BEV mask to the BEV-feature adversarial loss (ablation) | BEVCA · BEVFormer-base 3D detection | AP: 0.612 -> 0.191 | Improved |
| use BEV features instead of detection scores as attack loss (ablation) | BEVCA · BEVFormer-base 3D detection | AP: 0.564 -> 0.191 | Improved |
| use masked BEV loss instead of image-encoder feature loss (ablation) | BEVCA · BEVFormer-base 3D detection | AP: 0.392 -> 0.191 | Improved |
| use masked BEV loss instead of detection-head feature loss (ablation) | BEVCA · BEVFormer-base 3D detection | AP: 0.471 -> 0.191 | Improved |
| unmasked BEV loss over whole BEV map (ablation) | BEV-feature attack · BEVFormer-base 3D detection | AP: 0.618 normal -> 0.612 | Hurt |
| increase total variation weight beta (ablation) | BEVCA · BEVFormer series | attack performance | Improved |
| BEV mask thresholds delta r and delta theta set to 15 m and 15 degrees (ablation) | BEVCA · BEVFormer-base | AP | Improved |
| fine-tune victim models on BEVCA adversarial samples | BEV perception models · CARLA test | AP: 0.191 -> 0.615 | Improved |
| apply JPEG compression defense | BEVFormer-base under BEVCA · CARLA test | BEVCA AP: 0.191 -> 0.113 (QF=25) | No clear effect |
| heavy Gaussian input smoothing sigma=2.0 | BEVFormer-base under BEVCA · CARLA test | BEVCA AP: 0.191 -> 0.283 | Mixed |
| use BEVCA instead of CNCA in physical world | camouflage texture · Physical world, 4 BEV models | attack success rate: 30.00% -> 49.60% (BEVFormer base) | Improved |
Closest papers by meaning
- RAUCA: A Novel Physical Adversarial Attack on Vehicle Detectors via Robust and Accurate Camouflage Generation (ICML 2024 · similarity 0.88)
- M-BEV: Masked BEV Perception for Robust Autonomous Driving (AAAI 2024 · similarity 0.87)
- CNCA: Toward Customizable and Natural Generation of Adversarial Camouflage for Vehicle Detectors (NeurIPS 2024 · similarity 0.87)
- A Black-Box Evaluation Framework for Semantic Robustness in Bird’s Eye View Detection (AAAI 2025 · similarity 0.87)
- PhyCamo: A Robust Physical Camouflage via Contrastive Learning for Multi-View Physical Adversarial Attack (AAAI 2025 · similarity 0.86)
- Physical 3D Adversarial Attacks against Monocular Depth Estimation in Autonomous Driving (CVPR 2024 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
