An adversarial point cloud generation method, storage medium and terminal

Abstract

The invention discloses a method for generating an adversarial point cloud, a storage medium and a terminal. The method includes the following steps: destroying the topological structure of an input point cloud to generate an adversarial point cloud. The present invention adopts the method of changing and destroying the topological structure of the input point cloud, and then forms holes on the surface of the point cloud to generate the confrontation point cloud; selectively destroys the topological structure of the point cloud and discards the key topology of the point cloud, which will Greatly reduce the recognition accuracy of the model, and mislead the model to output wrong results; this method is superior to existing adversarial attacks in reducing the recognition accuracy of the model.

Description

technical field [0001] The invention relates to the field of adversarial point cloud generation, in particular to an adversarial point cloud generation method, a storage medium and a terminal. Background technique [0002] Point cloud is an important form of 3D data representation, which can be obtained directly from 3D acquisition devices. In recent years, deep neural networks (DNNs) have been successfully applied to point cloud data processing, the most popular of which are PointNet and its variant PointNet++. They extract features from raw point clouds in an end-to-end manner and have been applied in classification, segmentation, detection, and tracking scenes, achieving good performance. Recent studies have found that DNNs are vulnerable to adversarial attacks that can mislead deep networks by making imperceptible modifications to the original data. Existing spoofing attack algorithms mainly focus on the 2D attack domain. Recent studies have found that 3D point cloud ...

AI Technical Summary

Problems solved by technology

However, existing adversarial attack methods ignore the topology of point clouds

Images