Research On Multi-layer Neural Network Model Updating Facing Network Traffic

Multi-layer neural network is a method of deep learning.It is an end-to-end networked structure that can automatically perform features extraction,feature classification and other steps.Due to the automatic and precise characteristics of multilayer neural network,it is widely used in the field of network traffic classification.With the advancement of science and technology,the types and quantities of network traffic have increased rapidly.In order to ensure the effect of network traffic classification,the neural network model will not only develop toward a "wide and high" shape,but also need to use a larger-scale sample to train it,so that the neural network model update process will be very time-consuming.Therefore,it is necessary to optimize the model update process of the neural network to minimize the model update time while ensuring the recognition rate after the neural network model is updated.Computing offloading is an emerging technology applied in mobile cloud computing.The main idea is to select an appropriate offloading plan to offload tasks to cloud servers for execution based on factors such as time delay and energy consumption.In this way,the problems of weak computing power of mobile terminals unable to handle complex tasks and energy consumption of mobile terminals can be solved.The idea of computing offloading can be integrated into other scenarios,such as neural network model updates.This subject integrates computational offloading into the model update of the neural network,and then proposes the following update plan: compare different layers of the neural network to wireless terminals in the computational offloading scenario,and offload the update tasks of some layers during model update,do not update the parameters of these layers,only update the parameters of the remaining layers,in this way to reduce the time consumed in the model update process.More specifically,it is possible to combine the research on incremental learning and migration learning to divide the neural network into a certain layer,use this layer as the offloading layer,and offload all the update tasks on the layer without updating these layers.Only the parameters below the layer are updated.In this way,the model update time can be greatly reduced while ensuring the model recognition rate,this is the neural network model update scheme based on computational offloading proposed in this topic.Since this scheme is a theoretical scheme,and different neural networks have different characteristics,it is necessary to conduct experiments on specific neural network models.This topic uses convolutional neural network and capsule network to verify the above scheme,and uses the scheme proposed in this topic to update the model of convolutional neural network and capsule network.The experimental results obtained all prove the correctness of the proposed scheme.The conjecture of "the best offloading layer may have common features" was found when exploring the best offloading layer for further discussion.