Research On Fuzz Testing For End-to-End Autonomous Driving System

Autonomous driving systems can effectively control the driving behavior of the car and reduce the driving burden of the driver.It can be divided into multi-module autonomous driving systems and end-to-end autonomous driving systems.End-to-end autonomous driving system(EADS)has low construction cost.Therefore,it has been widely studied and applied by scholars.In order to ensure traffic safety,the EADS must be fully tested.Fuzz testing based on neuron coverage-guided is an effective testing method for EADS.But there are still some problems that need to be solved.Firstly,the existing fuzz testing based on neuron coverage-guided takes all decision logic of neural network as the test target,resulting in a long time to cover the main decision logic,which reduces the test efficiency.Secondly,fuzz testing based on neuron coverage-guided uses metamorphic relations to alleviate the problem of test oracle.However,the existing metamorphic relations of autonomous driving test scene are mainly constructed based on static environmental factors and meteorological factors,which still has shortcomings in guaranteeing the safety of the EADS in the scene containing dynamic environmental factors.In order to solve the above problems,this thesis focuses on improving the efficiency and adequacy of fuzz testing based on neuron coverageguided.The main research work of this thesis is as follows:(1)In view of the low test efficiency caused by the existing fuzz testing method based on neuron coverage-guided covering all neurons,this thesis defines the critical neurons and proposes a fuzz testing method based on the critical neuron coverage-guided,Critical Fuzz.The experimental results show that the critical neurons contain the main logic of the neural network.Compared with other fuzz testing methods,the average number of errors detected by the Critical Fuzz method is 4008,which is 5 times that of Deep Hunter and 4 times that of Tensor Fuzz.In addition,the average number of seeds used by Critical Fuzz is 78,0.5% of those used by Deep Hunter and 1% of those used by Tensor Fuzz.Therefore,Critical Fuzz is more effective and efficient than other fuzz testing methods.(2)In view of the single type of metamorphic relations constructed for automatic driving test scenarios by the existing fuzz testing method based on neuron coverage-guided,this thesis proposes a metamorphic relation based on the semantic of dynamic environmental factors.Because of the importance of traffic lights,this thesis constructs the semantic metamorphic relation of traffic lights.The experimental results show that the semantic metamorphic relation of traffic lights found that the traffic light recognition model has the problem of robustness.The recognition accuracy of traffic light recognition model maximum decreases by 58.20% when traffic lights are different states with the same semantics.The metamorphic relations of semantic change found that the traffic light recognition model has an average of 7,714 errors,and the EADS has an average of 74 errors.