PCV: A Point Cloud-Based Network Verifier

  • 2023-01-30 16:07:57
  • Arup Kumar Sarker, Farzana Yasmin Ahmad, Matthew B. Dwyer
  • 0


3D vision with real-time LiDAR-based point cloud data became a vital part ofautonomous system research, especially perception and prediction modules usefor object classification, segmentation, and detection. Despite their success,point cloud-based network models are vulnerable to multiple adversarialattacks, where the certain factor of changes in the validation set causessignificant performance drop in well-trained networks. Most of the existingverifiers work perfectly on 2D convolution. Due to complex architecture,dimension of hyper-parameter, and 3D convolution, no verifiers can perform thebasic layer-wise verification. It is difficult to conclude the robustness of a3D vision model without performing the verification. Because there will bealways corner cases and adversarial input that can compromise the model'seffectiveness. In this project, we describe a point cloud-based network verifier thatsuccessfully deals state of the art 3D classifier PointNet verifies therobustness by generating adversarial inputs. We have used extracted propertiesfrom the trained PointNet and changed certain factors for perturbation input.We calculate the impact on model accuracy versus property factor and can testPointNet network's robustness against a small collection of perturbing inputstates resulting from adversarial attacks like the suggested hybrid reversesigned attack. The experimental results reveal that the resilience property ofPointNet is affected by our hybrid reverse signed perturbation strategy


Quick Read (beta)

loading the full paper ...