| With the rapid development of big data and Internet of Things technologies,object detection technology is increasingly important in areas such as autonomous driving,intelligent security and scene recognition.However,in order to enhance the performance of the model,the traditional object detection technology needs to collect data from the client and upload it to the manufacturer’s server for centralized data annotation and model enhancement.However,this method has a high risk of data leakage.With the continuous improvement of data security protection laws,this behavior is subject to more and more constraints.Federated learning is a secure distributed machine learning technology proposed by Google,which attempts to fundamentally solve the problem of privacy leakage through the two principles of "data is not local to the user" and "gradient encryption".Based on federated learning technology,this paper designs a semi-supervised object detection system for automatic driving scenarios.The key problems to be solved are as follows:First of all,federal learning requires training to be completed locally by users,which puts forward high requirements on the lightweight of the federal learning framework.Excessive computing burden will make users resist to participate in the federal learning computing task,thus affecting the initiation of the federal learning task.To solve this problem,this paper designs a lightweight federated learning and training framework,defines a network communication protocol working in the transport layer,maximizes the communication efficiency between the server and the client,uses I/O reuse technology to improve the response speed of the server and reduce resource consumption.Secondly,the existing federal learning framework mostly uses differential privacy,secure multi-party computing,homomorphic encryption,these three gradient encryption algorithms,which all bring strong side effects to the training,either leading to the loss of model accuracy,or causing a great burden of encryption computing,so that"performance" and "security" pay attention to one and lose the other.To solve this problem,this paper proposes a noise self-elimination algorithm based on differential privacy.Based on the noise disturbance strategy based on differential privacy,the training client is grouped in pairs,the control nodes of each group are selected through the control node selection algorithm,and the control nodes of each group generate two groups of noise that can cancel each other and issue to the other member.Each group of clients can use the generated noise to confuse and encrypt the gradient information and then upload it to the aggregation server.By using noise cancelability,the aggregator server realizes automatic noise cancellation and completes gradient information aggregation in secret state.Compared with the difference privacy method with low accuracy and the secure multiparty calculation method with high computational burden,the proposed algorithm realizes the compact calculation and strong security,and avoids the negative impact on the model accuracy while resisting the gradient leakage attack.Finally,since local data is usually unlabeled original data automatically generated by terminal devices,it is impossible to realize effective utilization of these data in actual scenarios,so it is urgent to design an effective semi-supervised algorithm.In addition,the large difference in image style between each training terminal will also lead to the reduction of the cross-domain performance of the model,that is,the trained object detection model is only applicable to specific scenes,and the performance is not good in other scenes.Therefore,the precision reduction caused by cross-domain problems should be considered.To solve this problem,this paper proposes a semi-supervised object detection algorithm based on domain self-adaptation.It adopts the semi-supervised algorithm based on the average teacher model.Before the training task starts,part of public data set with high quality label is sent to the client,and teacher model and student model are trained at the same time.Guide the student model to learn the features exclusive to the client from the original data and public data set generated by the device,and complete the utilization of unmarked data;Through the network generated by CUT image style conversion,the style of training domain and target domain can be unified,so as to cope with the scene differences in different devices,to achieve data enhancement and improve the robustness of the model.A large number of experiments show that the proposed system can realize the secure federalization of the object detection algorithm and does not have a significant negative effect on the training speed.The proposed semi-supervised algorithm can also make good use of the unlabeled data locally,and effectively improve the accuracy and robustness of the model. |
