Research On IoT Device Identification Method Based On Network Traffic

The number of Internet of things(Io T)devices is growing rapidly.It has been widely used in smart home,transportation,medical care,and other fields,and have become an indispensable part of our life.However,many Io T devices exhibit security flaws making them vulnerable for attacks.And due to the manufacturers have difficulties in providing appropriate security patches to their products in a timely manner,Io T devices often accompanied by huge security risks.These devices are vulnerable to external hacker attacks thus cause information leakage,which will threaten the security of the network they connected.An effective way to solve this problem is to detect the suspicious Io T devices connected to the network,and then isolate them to avoid communication with the gateway.Therefore,the identification of Io T devices is of great significance to network management and network security.The research work of this paper is as follows.First,aiming at the identification of Io T devices,this paper designs an identification method of Io T devices based on convolutional neural network,which avoids the overhead of feature engineering and the possible errors during feature engineering.Experiments show that the classification accuracy of network traffic after data preprocessing is as high as 98%.Second,since deep learning models generally have large model sizes and slow computation speed,it is difficult to deploy these deep learning models in Io T devices with limited computing power and storage space.In this paper,a neuron pruning method based on compressed sensing is designed to lightweight the convolutional neural network and achieve the purpose of model compression.The experimental results show that the proposed pruning method can effectively reduce the model size and accelerate the calculation,but the performance loss is very small.Finally,because the traditional deep learning model concentrates all source data on a high computing power machine for training,this method leads to the unreliability of data flow and cannot ensure the safety of data and the training time is very long as well.This paper designs a federated learning model in which clients jointly train under the coordination of a central server.Experiments show that the federated learning model proposed in this paper can well protect the privacy and security of data,and can improve communication efficiency,reduce communication loss,and speed up training with little loss of classification performance.