MirrorAttack: Backdoor Attack on 3D Point Cloud with a Distorting Mirror

arXiv:2403.05847v1 Announce Type: cross
Abstract: The widespread deployment of Deep Neural Networks (DNNs) for 3D point cloud processing starkly contrasts with their susceptibility to security breaches, notably backdoor attacks. These attacks hijack DNNs during training, embedding triggers in the data that, once activated, cause the network to make predetermined errors while maintaining normal performance on unaltered data. This vulnerability poses significant risks, especially given the insufficient research on robust defense mechanisms for 3D point cloud networks against such sophisticated threats. …

abstract arxiv attacks backdoor breaches cloud cs.cr cs.cv data deployment embedding network networks neural networks processing security security breaches training type