Research On Generation Algorithm Of Laser Point Cloud Adversarial Sample Under Automatic Driving Situation

Due to the continuous progress and development of science and technology,automatic driving technology has received more and more research and attention,and the security issues in driverless,especially the security issues at the perception level are important.Research shows that the target detection model in automatic driving vehicles is fragile and will be affected by specific interference,which is often imperceptible to people.Li DAR is a kind of sensor that has been favored by automatic driving applications recently.It can obtain three-dimensional images of the environment,thus helping the system to judge the depth information conveniently.However,the research on the vulnerability of target detection model mainly focuses on visual sensors,and the research on the security of target detection model based on Li DAR is still less.In this paper,the vulnerability mechanism of laser point cloud target detection is studied,and the generation method of adversarial examples in the automatic driving situation is proposed.The specific research contents are as follows:(1)A new point cloud adversarial method based on moving point and local saliency is proposed,which takes a single target dense point cloud as the target.This method is different from the previous methods to attack the global point cloud,but selects the local point cloud with significance based on the gradient to attack,thus greatly improving the imperceptibility of the adversarial examples while maximizing the attack performance.At the same time,this method proposes two intra-class shrinkage losses based on the result preference,which makes the iterative process of training adversarial examples converge faster and has higher performance attack them,and proposes a Perturbationbudget Withdrawing Algorithm(PWA)to release unnecessary distortion budget,so that the optimization process can get rid of local solutions.(2)A differentiable mapping method that can map the target point cloud to the real environment point cloud is proposed,so that the iterative adversarial examples can be optimized in the driverless scenario.On this basis,a point cloud adversarial examples generation algorithm for the real driverless situation is proposed.The algorithm uses the targeted loss function to train the adversarial examples,so that the final generated adversarial examples cannot be recognized by the point cloud target detector.At the same time,an optimization algorithm based on Fast Point Feature Histograms(FPFH)is proposed,which can effectively reduce the number of outliers and improve its imperceptibility.(3)Based on the proposed point cloud adversarial examples generation algorithm under unmanned driving scenario,a point cloud adversarial examples generation system with high availability and high stability is constructed.It is convenient for users to select arbitrary data sets and target models,and use the system to quickly and directly customize and generate point cloud adversarial examples.The experimental results show that the point cloud adversarial examples generation method proposed in this paper has higher attack performance and stronger insensibility,and its adversarial examples can be applied to the real automatic driving scene through 3D printing and other ways.