TY - JOUR

T1 - 3DVerifier: efficient robustness verification for 3D point cloud models

AU - Mu, Ronghui

AU - Ruan, Wenjie

AU - Soriano Marcolino, Leandro

AU - Ni, Qiang

PY - 2022/11/2

Y1 - 2022/11/2

N2 - 3D point cloud models are widely applied in safety-critical scenes, which delivers an urgentneed to obtain more solid proofs to verify the robustness of models. Existing verifcationmethod for point cloud model is time-expensive and computationally unattainable on largenetworks. Additionally, they cannot handle the complete PointNet model with joint alignment network that contains multiplication layers, which efectively boosts the performanceof 3D models. This motivates us to design a more efcient and general framework to verifyvarious architectures of point cloud models. The key challenges in verifying the large-scalecomplete PointNet models are addressed as dealing with the cross-non-linearity operationsin the multiplication layers and the high computational complexity of high-dimensionalpoint cloud inputs and added layers. Thus, we propose an efcient verifcation framework,3DVerifer, to tackle both challenges by adopting a linear relaxation function to bound themultiplication layer and combining forward and backward propagation to compute the certifed bounds of the outputs of the point cloud models. Our comprehensive experimentsdemonstrate that 3DVerifer outperforms existing verifcation algorithms for 3D models interms of both efciency and accuracy. Notably, our approach achieves an orders-of-magnitude improvement in verifcation efciency for the large network, and the obtained certifedbounds are also signifcantly tighter than the state-of-the-art verifers. We release our tool3DVerifer via https://github.com/TrustAI/3DVerifer for use by the community.

AB - 3D point cloud models are widely applied in safety-critical scenes, which delivers an urgentneed to obtain more solid proofs to verify the robustness of models. Existing verifcationmethod for point cloud model is time-expensive and computationally unattainable on largenetworks. Additionally, they cannot handle the complete PointNet model with joint alignment network that contains multiplication layers, which efectively boosts the performanceof 3D models. This motivates us to design a more efcient and general framework to verifyvarious architectures of point cloud models. The key challenges in verifying the large-scalecomplete PointNet models are addressed as dealing with the cross-non-linearity operationsin the multiplication layers and the high computational complexity of high-dimensionalpoint cloud inputs and added layers. Thus, we propose an efcient verifcation framework,3DVerifer, to tackle both challenges by adopting a linear relaxation function to bound themultiplication layer and combining forward and backward propagation to compute the certifed bounds of the outputs of the point cloud models. Our comprehensive experimentsdemonstrate that 3DVerifer outperforms existing verifcation algorithms for 3D models interms of both efciency and accuracy. Notably, our approach achieves an orders-of-magnitude improvement in verifcation efciency for the large network, and the obtained certifedbounds are also signifcantly tighter than the state-of-the-art verifers. We release our tool3DVerifer via https://github.com/TrustAI/3DVerifer for use by the community.

U2 - 10.1007/s10994-022-06235-3

DO - 10.1007/s10994-022-06235-3

M3 - Journal article

VL - 2022

JO - Machine Learning

JF - Machine Learning

SN - 0885-6125

ER -