Research On Neural Network Algorithm Based On Vision Sensor And Safety In Autonomous Driving

With the maturation of Artificial Intelligence-related technologies and the continuous development of the intelligent service industry,deep neural networks have been widely used in fields such as autonomous vehicles and object detection.However,research in recent years has shown that deep neural network models are susceptible to deception by adversarial perturbation images,leading to detection failure or misclassification of object detection models(e.g.,traffic sign detection models,pedestrian detection models,and obstacle detection models)applied to autonomous vehicles.The safety hazards of object detection models can cause decision errors in autonomous systems,which can lead to the loss of human and property.Therefore,in order to further verify the existence of security vulnerabilities in deep neural networks,this paper conducts research from two perspectives: based on the physical world adversarial and based on the digital world adversarial.In the study based on existing physical world adversarial method,when the adversarial object is a traffic sign,the adversarial perturbed image generated by the adversarial method is vulnerable to environmental factors.In the study based on existing digital world adversarial method,the adversarial method suffers from the problem of gradient dispersion and does not utilize the feature information provided by the convolutional layer in the object network model during training.To solve the above problems,the main work of this paper is as follows:(1)Aiming at the problem that the adversarial perturbation images of traffic signs are vulnerable to environmental factors,this paper proposes an adversarial perturbation image generation method applied to a traffic sign detection model.The method introduces a loss function to constrain the expected effective distance error between the input of the adversarial perturbation image and the input of the original image in the total loss function of the generated adversarial perturbation image to improve the adversarial effect of the adversarial perturbation.A multi-class data enhancement method is proposed to expand the adversarial perturbation image set by affine transformation and appearance transformation to improve the robust performance of the adversarial perturbation images when training the adversarial perturbation images.Experimental results show that the adversarial perturbation images generated by this method reduce the m AP value of the traffic sign detection model to 25.6%,and the overall adversarial performance is improved compared with existing physical world-based adversarial methods.(2)In response to the problem of Gradient vanishing and not using convolutional layer features in existing digital world adversarial methods,this paper applys WGAN-GP to adversarial perturbation image generation method.The method adopts the WGAN-GP objective loss function instead of the traditional objective loss function to solve the problem of the existence of gradient vanishing in the generator.And a feature extractor is introduced in the network framework to convert the GAN network generator from learning the features of the original image to learning the object features,which enhances the concealment of the counteracting perturbation image and improves the counteracting effect.Experimental results based on different object network models show that the adversarial perturbation images generated by this method outperform the adversarial perturbation images generated by existing digital world-based adversarial methods in terms of adversarial performance and image quality.(3)In order to visualize and facilitate users to selectively generate the adversarial perturbation images generated by the above two adversarial methods,this paper designs and implements an adversarial perturbation image generation system.The architecture of the system is designed to effectively reduce the coupling between modules and improve the scalability.The system provides an experimental test platform for subsequent research work.