MiuGS: Modality-Integrated Unified Gaussian Representation for Physical Adversarial Object Generation in 3D Detection

Abstract

Multimodal 3D object detection is a cornerstone of autonomous driving perception, yet its adversarial robustness remains poorly understood. Due to the lack of a unified cross-modal representation, existing adversarial attacks manipulate image and point cloud inputs separately, leading to limited physical realizability and suboptimal attack effectiveness. To address this problem, we propose MiuGS, a modality-integrated unified Gaussian representation framework for generating 3D adversarial objects with strong attack effectiveness and physical-world potential. For physical realizability, we design a multimodal 3D Gaussian coordination strategy to address the imprecise cross-modal fusion issue. Specifically, MiuGS models the adversarial object as a shared 3D Gaussian representation and instantiates it into consistent image and point cloud observations, enabling the two sensing streams to be coupled at the representation level rather than optimized independently. To further strengthen attack effectiveness, we introduce a cross-scenario guidance rebalancing strategy to mitigate cross-modal guidance drift across scenarios and the digital-to-physical domain gap. Extensive experiments show that MiuGS achieves strong attack effectiveness, reaching an average ASR of 78.8% on nuScenes and 44.3% on CARLA, supporting the effectiveness and physical-world potential of the proposed unified representation.