Research On Privacy-preserving Object Detection In Connected Autonomous Vehicles

Connected autonomous vehicles(CAVs)adopt rich sensors to perceive environment,combine the deep neural network model to detect the object in the field of vision,then apply the feedback detection results to adjust the driving strategy.Considering the limitations of computing resources,storage resources and endurance,CAVs outsource the neural network inference tasks to the edge nodes.However,raw data contains a large amount of private information,and edge nodes are semi-trusted entity,which means that CAVs uploading raw data can lead to serious privacy leakage issues.In this dissertation,the key technologies of privacy computing for CAVs are researched from four aspects:secure computing protocols designing,image-oried object classification,image-oried object detection and point cloud-oried object detection.we explore the universal design idea of secure computing protocols and construct edge-assisted privacy-preserving framework,aiming at reconciling the contradiction among security,correctness and efficiency.The research provide theoretical and technical support for secure data sharing among CAVs.The main research works are as follows.1.Aiming at the computing privacy disclosure of encrypted deep neural networks,this dissertation proposes a universal and efficient protocol design method for secure two-party computing based on additive secret sharing technology.For linear operations such as addition,it can be directly split into two shares and calculated independently by both parties.For non-linear operations such as division,power,exponent and comparison,we design a set of secure transform protocols between additive sharing and multiplicative sharing,which can solve the problem of splitting nonlinear functions effectively by combining with the basic mathematical properties.Theoretical analysis ensures the correctness and security of proposed protocols,and which supports paralle computing.Experiments show that these protocols have efficient efficiency and performance.2.In order to solve the category privacy disclosure in the image-oried object classification,this dissertation proposes a privacy-preserving object classification framework.CAVs only need to randomly split and upload the image share,which reduces the local computational cost to the greatest extent.Based on secure computing protocols designed,two edge servers cooperate to perform computations involved in convolutional layer,activation layer,pooling layer and full-connected layer.Compared with primitives such as homomorphic encryption,our secure computing protocols has great efficiency advantages.The correctness and security of the classification framework are analyzed theoretically.The experimental results show that the performance of the framework is better than the existing work by using the KITTI dataset and VGG16 model.3.For the category and location privacy disclosure in the image-oried object detection,this dissertation proposes a privacy-preserving object detection framework,which can achieve the object feature extraction and location detection.A privacy-preserving region proposal network is proposed,which securely realizes the fitting process between anchor box and predicted object bounding-box,as well as object bounding-box correction,non-maximum suppression and other operations,finally obtaining the object location sharing.Furthermore,this dissertation designs a secure feature matching method,which extracts the feature map sharing of each object without revealing the real object location and feature information.Theoretical analysis shows that the proposed detection framework has the same complexity as the plaintext detection model,and the experimental results with KITTI dataset are consistent with the theoretical analysis.4.Compared with images,point clouds have a stronger ability to understand three-dimensional scenes.To adress the category and location privacy disclosure in the point cloud-based object detection,this dissertation proposes a point cloud-oried privacy-preserving object detection framework.First,two edge servers collaborate to generate 3D bounding boxes from the raw point cloud in a bottom-up manner.Then,the point features are converted into canonical coordinates,and the global semantic features and spatial pooling features are integrated to further standardize the bounding box and obtain the final object detection results.Through theoretical analysis and experiment,the proposed framework can ensure the correctness of the whole computing process,and effectively protect the location and classification privacy of the point cloud object.