Design Of Actuator Attack Detection Algorithm For Unmanned Driving Syste

In recent years,breakthroughs in artificial intelligence research have effectively promoted the development of driverless technology.Commercial companies and research institutes have carried out actual road tests of autonomous driving.However,the safety of autonomous vehicles must be addressed first before they enter the mass market.At present,the research on the safety of autonomous vehicles is mostly based on sensor security and network communication security.The actuator,as the last component of the interaction between autonomous vehicles and the external environment,is very important to ensure the safe driving of autonomous vehicles,but the safety of the actuator has not received corresponding attention.The thesis focuses on the problem of actuator attack detection for autonomous vehicles,and does the following work:(1)Actuator attack detection algorithm based on artificial threshold settingFirstly,according to the measurement data of vehicle sensors,combined with the maximum likelihood estimation method and the vehicle dynamics model,the actual control instructions of the autonomous vehicle are estimated.Then,the estimated actual control instructions are compared with the reference control instructions issued by the vehicle computer.If the difference is greater than a certain threshold,the actuator is considered to be attacked.On this basis,different attack types can be identified by setting different thresholds.Finally,the simulation experiments on the Gazebo and Autoware simulation platforms show that the method designed in this thesis has good effects in the detection of actuator attacks.(2)Actuator attack detection algorithm based on convolutional neural networks(CNN)Applying CNN to actuator attack detection can avoid setting threshold artificially and improve the adaptability of the algorithm.Firstly,the data of the driving state of the autonomous vehicles are collected,that is,the difference between the actual execution of the vehicle control command and the upstream controller sending command estimated in work(1).In order to use the image recognition function of CNN,the difference curve is transformed into an image,and the curve is manually marked whether it is in an abnormal state.As a training data set,it is sent to CNN for training.Finally,the trained convolutional neural network model is used for actuator attack detection.The test results show that the convolutional neural network model can improve the detection efficiency to a certain extent,and improve the adaptability of the detection algorithm.(3)Actuator attack detection algorithm based on long-short term memory network(LSTM)autoencoderLSTM autoencoder does not need to manually set data labels,which avoids tedious labeling work.Firstly,an autoencoder model with two long-short term memory networks is designed.Then the relevant data in the driving process are used as a training set to train the auto-encoder.Finally,the trained LSTM autoencoder is used in attack detection simulation.The test results show that,without manually annotating data,LSTM autoencoder can automatically learn to identify the abnormal behavior in the difference curve and effectively detect possible attacks.