Adversarial sample generation method and system for laser radar

Abstract

The invention discloses an adversarial sample generation method for a laser radar. The method comprises the following steps: 1) finding out a target needing to be attacked and a fragile region of the target in a point cloud picture; (2) generating independent points of antagonism in the target area in the step (1); (3) optimizing the antagonistic independent points generated in the step (2) based on a disturbance index; 4) on the basis of the step 3, generating antagonistic clusters based on the antagonistic points; and 5) optimizing the adversarial cluster generated in the step 4 based on a disturbance index, and finally generating an adversarial sample. The invention further provides a system for implementing the adversarial sample generation method for the laser radar. The system is formed by connecting a target detection and screening module, an adversarial point generation module, an adversarial point optimization module, an adversarial point cluster generation module and an adversarial cluster optimization module. The laser radar point cloud adversarial sample generated by the invention is high in confrontation, the loss is small relative to the original data, and the basic features of the original data can be better reserved.

Description

technical field [0001] The present invention relates to the field of deep learning security technology, specifically a method and system for generating adversarial samples for laser radar, which generates adversarial points by analyzing vulnerable areas of original point cloud targets, and optimizes the positions and shapes of adversarial points to generate adversarial clusters , and finally generate the adversarial samples and systems of the attack target. Background technique [0002] With the rapid development of artificial intelligence technology, driverless cars are turning from science fiction to reality. The autonomous driving technology of unmanned vehicles is mainly divided into three parts: perception, decision-making and control. The perception module provides an important basis for its decision-making and control. However, due to the limited perception capabilities of existing sensors, they are easily affected by external objective physical factors, which leads...

AI Technical Summary

Problems solved by technology

[0007] For the generation of adversarial samples for 2D images, there are currently many researches in the industry and there are more mature solutions, but for 3D images, especially the generation of adversarial samples for 3D point cloud images generated by lidar is relatively less related work

Images