The Design And Implementation Of Lidar Fuzzing System For Autonomous Driving Systems

Li DAR is an important environmental sensing sensor for high-level autonomous driving,which can generate three-dimensional point cloud data that is more advanta-geous than two-dimensional data in terms of information expression.With the increas-ing popularity of Li DAR in autonomous vehicles,the safety and stability of Li DAR have also received more and more attention.Especially in terms of autopilot software,because Li DAR uses deep learning technology for target detection,it is difficult for the input domain space to cover all possible scenarios,and there may be loopholes in security.Therefore,how to effectively test the Li DAR model of autonomous driving software has become a key point of research.Fuzzing is a widely used quality assurance and vulnerability discovery technique,which tests the target program by generating a large number of unexpected inputs.In the machine learning system,fuzzing can greatly improve the efficiency and coverage of the test,so it is feasible to use fuzzing to improve the adequacy of Li DAR test of au-topilot software.However,the data characteristics of point cloud data are quite different from those of pictures and text.Therefore,it is necessary to design some new fuzzing mutation operators for the point cloud data of autonomous driving Li DAR.In this the-sis,by analyzing the characteristics of the number,coordinates and intensity of the point cloud data of autonomous driving Li DAR,combined with the actual scene information,a series of fuzzing mutation operators for the point cloud data of autonomous driving Li DAR are designed.Subsequently,based on the Spring Boot framework,a Li DAR fuzzing system for autonomous driving software was implemented.Through this sys-tem,users can upload test data sets and test model,and then select mutation operators and configure test task parameters to execute the test tasks.The system will generate disturbance test cases based on the designed point cloud special mutation operator,and automatically deploy and test the test model,and generate test reports.In addition,the system also provides point cloud visualization functions,with strong usability.This thesis conducts functional tests on the system based on the system's demand analysis and outline design to ensure the usability of the system.Then,based on the Point Cloud Target Detection Network,POINTPILLARS,an empirical research on this system is carried out to further verify the effectiveness of the point cloud mutation operator and the Li DAR fuzzing for autopilot software proposed in this thesis.The experimental results show that the perturbation test cases generated by the point cloud special mutation operator in this paper are real and effective,and the fuzzing test method proposed in this paper can reveal the problems of the autonomous driving lidar model.The fuzzing system developed in this thesis can effectively test the autopilot software Li DAR,improve the adequacy of the test,and have certain positive significance for improving the safety of the autopilot software Li DAR.