摘要
多模态感知系统结合了来自摄像头、激光雷达、雷达和其他传感器的数据,对于驾驶自动化的可靠决策至关重要。然而,这些系统仍然容易受到对抗性攻击,特别是在数据模式不一致的情况下。现有的对抗性攻击方法难以在多模态、异质性强、实时性要求高的驾驶自动化场景中实现有效干扰。为此,本文提出一种跨模态一致的对抗性样本生成框架(CMAEG),融合跨模态一致性优化机制、多目标攻击损失函数设计与轻量化生成架构,实现对图像、LiDAR等多模态输入的统一攻击。CMAEG引入了一个跨模态一致性优化模块和一个多目标损失机制,在保证攻击成功率的同时,显著提升对抗样本的隐蔽性与跨模态协同一致性,并具备优良的实时性和泛化能力。此外,本文还为对抗性样本的实时生成设计了轻量级生成器体系结构。在驾驶自动化数据集(nuScenes,KITTI)和融合模型(例如BEVFusion,TransFusion)上进行的大量实验证明了本文方法在保证攻击成功率的同时可以实现较好的隐蔽性和实时性。
The multi-modal perception system,which integrates data from cameras,lidars,radars,and other sensors,is crucial for reliable decision-making in autonomous driving.However,these systems remain vulnerable to adversarial attacks,especially when the data patterns are inconsistent.Existing adversarial attack methods are difficult to effectively interfere in the autonomous driving scenarios with diverse modalities,strong heterogeneity,and high real-time requirements.To address this,this paper proposes a cross-modal consistent adversarial example generation framework(CMAEG),which integrates a cross-modal consistency optimization mechanism,a multi-objective attack loss function design,and a lightweight generation architecture to achieve unified attacks on multi-modal inputs such as images and LiDAR.CMAEG introduces a cross-modal consistency optimization module and a multi-objective loss mechanism,significantly enhancing the concealment and cross-modal collaborative consistency of adversarial samples while ensuring the attack success rate,and it also possesses excellent real-time performance and generalization ability.Additionally,we have designed a lightweight generator architecture for the real time generation of adversarial samples.Extensive experiments on autonomous driving datasets(nuScenes,KITTI)and fusion models(e.g.,BEVFusion,TransFusion)demonstrate that our method can achieve good concealment and real-time performance while ensuring the attack success rate.
作者
胡佳俊
薛吟兴
商广旭
Hu Jiajun;Xue Yinxing;Shang Guangxu(University of Science&Technology of China,Hefei 230000)
基金
汽车标准化公益性开放课题(编号CATARC-Z-2024-01127)。
关键词
自动驾驶
对抗性攻击
跨模态一致性
联合损失优化
autonomous driving
adversarial attack
cross-modal consistency
joint loss optimization
