Research On Intrusion Detection Technology Based On Combined Neural Network For Internet Of Things

With the rapid development of the information age,the amount of network data and the amount of Internet of Things equipment connections are becoming larger and larger.Due to factors of the Internet of Things such as scalability,resource constraints,large number of connected devices and inconsistent security standards,Internet of Things equipment and systems are more vulnerable to intrusion,and collapse and paralysis under the situation that complex network attack methods emerge in an endless stream.Since the traffic data has the characteristics of space and time series dimensions,and the combined neural network can extract features of different dimensions in a targeted manner,this paper aims at the problem of deep spatiotemporal feature extraction of one-dimensional traffic data,using the combined neural network structure to improve,and combined with the improved data balance algorithm to optimize the data imbalance problem in the data set.The details of main works are as follows:(1)Aiming at the spatiotemporal feature extraction of traffic data,this paper designs a one-dimensional densely connected convolutional neural network named Dense Net1 D,which reuses the underlying features through a one-dimensional densely connected structure to extract deep spatial feature information of the data;at the same time,the time series characteristic relationship of the data is considered,adding GRU unit to further extract the potential time series features of the data,through this improved combined neural network structure to fully learn the deep space and time series features of the traffic data,to achieve the purpose of global deep spatiotemporal feature extraction.(2)The data imbalance problem in the data set will affect the detection rate of some minority samples and the overall decision-making of the model.To solve this problem,this paper proposes an improved data balance method to optimize it based on the combined neural network model.At the data level,the ADASYN oversampling algorithm is used to generate the one-dimensional traffic data of the minority class,and then the NCL downsampling algorithm is used to eliminate a certain amount of the majority class data that causes classification difficulties;at the algorithm level,the focal loss function with self-defined weight is used to further strengthens the balance of the overall data,thereby improving the overall detection effect of the model and the detection rate of minority samples.(3)Based on the above mentioned methods,this paper selects the network intrusion detection data set UNSW-NB15,which extracts the self-built data subset NB15-Io T that reflects the characteristics of the Internet of Things and common Internet of Things attack traffic,and the professional Internet of Things data set Bot-Io T to carry out multi-classification experiments respectively.Through the comparative analysis of experiments,the detection effect is improved compared with other methods.At the same time,the Internet of Things intrusion detection system is designed based on this method,which verifies the effectiveness and practicability of this research.