Research And Implementation Of Adversarial Example Generation Of Point Cloud

Deep neural networks have demonstrated impressive success in perception tasks such as autonomous driving,robotics,and virtual reality.However,these networks are vulnerable to adversarial attacks,which can cause the target system to behave abnormally.Adversarial attacks on deep neural networks mainly fall into two categories:adversarial sample attacks and neural network backdoor attacks.With the combination of deep learning and 3D sensors such as Li DAR,3D point cloud deep learning is being increasingly applied in various safety-critical applications.Unfortunately,point cloud neural networks are also severely threatened by adversarial attacks,which increase the risk in safety-critical scenarios and may lead to serious security accidents.In this thesis,we first propose a method for generating adversarial examples for 3D point clouds quickly.The method includes a self-organizing map encoding network and a dual-stream decoding network.The self-organizing map encoding network uses a self-organizing map to group the original point clouds,extract global features,and merge them into a vector to encode the point clouds into global features quickly and efficiently.The dual-stream decoding network uses convolutional and fully connected branches to generate adversarial examples and merges them to generate the final adversarial example.This method can quickly restore the global features of the point clouds into 3D point clouds and introduce adversarial loss during the restoration process,making the restored 3D point clouds capable of deceiving the classifier.Moreover,this method is almost independent of the victim network,so it can be more widely applied to different networks.This thesis proposes a point cloud backdoor attack method aimed at attacking point cloud classification networks.Similar to backdoor attacks on image classification networks,this attack method aims to make the 3D point cloud classifier learn to misclassify trigger-embedded samples as the target class set by the attacker during training.When testing data contains trigger-embedded samples,the network will misclassify them as the class set by the attacker.This article’s method optimizes the shape point cloud density and spatial position of the embedded trigger to achieve better attack performance.This thesis investigates the security issues of point cloud deep neural networks,mainly focusing on adversarial examples and backdoor attacks in neural networks,as well as the system deployment and implementation of adversarial example generation algorithms.This thesis provides significant reference value for the generation of point cloud adversarial examples and point cloud backdoor attacks,as well as the practical use and system deployment of point cloud adversarial examples.In summary,this thesis primarily investigates the security issues of point cloud deep neural networks,including adversarial examples and backdoor attacks,as well as the system deployment and implementation of point cloud adversarial example generation algorithms.The research findings of this thesis have significant reference value for the practical use and system deployment of point cloud adversarial examples.