Safety Verification Of Steering Angle In Autonomous Driving

The rapid development of deep learning technology has made a qualitative breakthrough in the research and deployment of DNN-driven autonomous driving technology.However,while autonomous vehicles have driven tens of billions of kilometers on the road without humans in operation,the safety of their autonomous behavior can’t be guaranteed —— even small deviations in factors such as steering angle,safe distance,acceleration and braking can have unpredictable consequences for safe manoeuvring.Therefore,the safety verification technology of autonomous driving vehicles has gradually become the current research hotspot and difficulty.In this paper,the image data collected by the camera sensor is taken as the input and the correct steering angle is taken as the output to study the safety verification of the steering angle of DNN-driven autonomous driving vehicles.At present,the representative methods of DNN security verification mainly include Deeptest,DLV and SDLV.DeepTest automatically synthesizes test cases through multiple image transformations.Based on the test cases,it automatically detects mistakes made by DNNdriven vehicles that could lead to fatal collisions.However,the framework is limited by the types of image transformation and cannot simulate all kinds of driving scenes in the real world.DLV is an automatic verification framework for security verification of image classification neural networks.Based on the satisfiability mode theory,the framework performs security verification of image classification decisions related to image operation through layer by layer disturbance analysis of neural networks.Since the correct steering command of the autopilot system is not unique,the DLV algorithm cannot be directly used to verify the safety of the steering angle of the autopilot system.SDLV is an automatic verification technology for the safety verification of steering angle of autonomous driving vehicles.This technique extends the DLV verification framework by using the coverage of neurons and the relaxation relation to solve the judgment problem of the prediction behavior,so as to realize the safety verification of the steering angle.The CSDLV verification framework proposed in this paper is further optimized and improved against SDLV framework.On the one hand,safety verification is realized by finding antagonistic misclassification.Therefore,in order to further improve the success rate of finding antagonistic cases,this paper redefines the neuron coverage in the deterioration relationship,and the main functional area neuron coverage and strong neuron coverage are introduced to extend the DLV framework.On the other hand,there are false positive samples in the adversarial cases found by SDLV.In order to reduce the proportion of false positive in the adversarial misclassification,this paper adopts a combination of relaxation relation and filtering criteria to further screen the adversarial cases.This paper evaluates the proposed automated verification framework on NVIDIA’s end-to-end autonomous driving system,using a data set containing image data from vehicle dashboard cameras marked with steering angles collected near California and San Pedro,USA.The final experimental results show that this automatic verification framework can successfully find the antagonistic misclassification.Meanwhile,compared with SDLV,the success rate of CSDLV framework to find the antagonistic cases increases from 71.8% to 79.6% under the current optimal parameter configuration,and the proportion of false positive samples in the experimental results decreases from 5.014% to3.517%.