BEVCA: Effective and Transferable Camouflage Attack against Multi-View 3D Perception in Autonomous Driving

Linye Lyu, Jiawei Zhou, Daojing He, YU LI

Keywords: Adversarial Attack; Vehicle Detection; Vision-based Perception; Autonomous Driving;

TL;DR: BEVCA: the first framework to generate adversarial camouflage that effectively attacks multi-view 3D perception models

Abstract: Multi-view 3D perception models are widely adopted by leading car manufacturers due to their highly competitive performance. However, existing adversarial camouflage techniques primarily focus on single-view 2D detectors, limiting their effectiveness against multi-view 3D perception models. In the paper, we propose BEVCA, the first framework to generate adversarial camouflage that effectively attacks multi-view 3D perception models by exploiting the Bird's-Eye-View (BEV) representation used across various 3D perception models. Our framework introduces a new differentiable multi-view neural renderer to enable end-to-end gradient-based camouflage optimization. Furthermore, we propose a novel BEV-feature-based adversarial loss to achieve effective and transferable attacks. Extensive experiments on 3D object detection and segmentation scenarios demonstrate that BEVCA outperforms existing baselines, achieving attack improvements of 36.2\% and 21.6\% in black-box settings, respectively.

Primary Area: alignment, fairness, safety, privacy, and societal considerations

Submission Number: 8685