Research On CNN Model Training And Inference Technology For Edge Intelligence

With the development of artificial intelligence(AI)and the Internet of Things,more and more terminal devices have edge computing capability,which has been widely used in intelligent transportation,intelligent industry and other fields.As a result,it has been a rapid development in the direction of edge intelligence(EI),which is a way of applying AI to edge computing.EI is a series of intelligent methods for training and infering machine learning models with the aid of edge side.With the advent of high performance system on Chip(So C),running convolutional neural network(CNN)applications on mobile devices has gained a great deal of attention.However,the computing capacity of a single terminal device is limited,and it is challenging to perform computationally intensive CNN model infering and training on the terminal device,resulting in high delay and energy consumption.To solve this problem,the scenario of edge computing usually adopts the way of computing task unloading,which can unload part of CNN model to edge nodes.Therefore,an efficient unloading method needs to be studied.In addition,this method will bring certain data security problems,such as data attacks,uninstallation data tampering,which resulting in inaccurate infering and training results.To solve the above problems,the corresponding research methods of end-to-end joint inference and joint training are respectively proposed in this thesis.The specific contents are as follows:(1)To solve the Multi-CNN(MCNN)model inference problem in edge computing environment,a MCNN infering method Deep INT based on data integrity was proposed.First,the MCNN inference process is modeled as a workflow model where the task represents a convolution or full join computation.The task execution workload is obtained by inferring the time cost of the task in the experiment.To ensure the integrity of the task,a complete data transmission mechanism is designed to minimize the transmission time during uninstallation.The integrity model is established based on the time cost of task integrity verification.On this basis,a Deep INT algorithm based on particle swarm optimization(PSO)is proposed to design multi-dimensional coding strategy by considering task unloading position,GPU frequency and CPU frequency.Finally,the experimental results show that Deep INT is significantly better than the existing system,greatly reducing the energy consumption and improving the model inference efficiency.(2)To solve the CNN model training problem of multiple heterogeneous terminal devices,a distributed joint learning method DJL is proposed,in which the edge server assists multiple terminal devices to train.In order to solve the contradiction between model unloading and data privacy protection of terminal devices,an optimal model unloading strategy is proposed first,which can improve the efficiency of model training as much as possible while protecting data privacy.In order to reduce the discretization effect caused by the heterogeneity of terminal devices,a device terminal model allocation strategy was proposed based on a large number of experiments and the analysis method of polynomial regression model,so as to optimize the balance between computational efficiency and performance.Finally,we implemented DJL on a test platform using multiple heterogeneous terminal devices.Experimental results show that DJL has advantages in training efficiency and privacy protection.In summary,the security and efficiency of the proposed method are demonstrated through data verification in two actual scenarios.The experimental results show that the method proposed in this thesis has better performance.