Algorithm Design And Implementation Of GNSS Anti-Spoofing For Autonomous Vehicle Based On Machine Learning

Autonomous vehicles need GNSS(Global Navigation Satellite System)to provide accurate,reliable and continuous real-time positioning information in order to implement its safety-critical applications such as autopilot and navigation.However,the GNSS signal is vulnerable to spoofing attacks.Malicious spoofer can manipulate the signal to deceive the receiver,causing the vehicle to deviate from its expected path or stop working properly.Even the previously believed reliable multi-sensor fusion(MultiSensor Fusion)positioning system has become insecure with the emergence of new spoofing attacks.Therefore,in order to resist complex and variable spoofing attacks and ensure the reliability and safety of autonomous vehicles,both industry and academia urgently need anti deception technologies with better performance and lower cost.In order to address the above challenges,this thesis proposes two GNSS anti spoofing methods based on machine learning.The main contributions of this thesis are as follows:Autonomous GNSS Anti-Spoofing Architecture:Based on the background that MSF positioning is no longer secure,this thesis makes a theoretical analysis of the newly emerged GNSS attacks,constructs a new GNSS attack model and attack scenarios,and proposes a GNSS anti-spoofing architecture based on this,which can realize customized training of GNSS anti-spoofing models through real-time interaction between digital twins and vehicles,while reducing the burden on vehicle hardware.GNSS anti-spoofing method based on gradient boosting tree:In this thesis,a prediction-based supervised learning approach is proposed for spoofing detection.The method directly learns the mapping relationship between real-time vehicle sensor data and the actual displacement through the gradient boosting regression tree,which avoids the error problem in MSF positioning systems and greatly improves the accuracy of position prediction.The method also uses a gradient boosting classification tree to learn a reasonable deviation between the predicted position and the localization output,and performs spoof detection and localization recovery based on this deviation.GNSS anti-spoofing method based on generative adversarial networks:This thesis proposes a generative unsupervised learning method for spoof detection.The generative adversarial network can generate anomalous data and capture unknown attacks,in which the generator can simulate the localization timing data,which well solves the problem of insufficient samples of anomalous data; the encoder can map the timing data to the implicit layer,which is inter-invertible with the generator to reconstruct the multidimensional timing data of the vehicle sensors; the discriminator can output the true probability of the joint distribution of the localization timing data and the implicit layer representation.The designed anomaly detection algorithm combines the feature matching error and reconstruction error of the discriminator to calculate the anomaly score,and uses the dynamic thresholding technique to determine the deception.The experiments show that the proposed two anti-spoofing methods both demonstrate greatly improved F1-score compared with other methods,and can well resist new types of attacks against the automated MSF positioning system.They are very good in dealing with both known and unknown attacks without additional hardware costs,and offer broad application prospects.