Intelligent Detection Technology Of IoT Botnet Traffic Against Concept Drift

The rapid development of the Internet of Things(Io T)has brought many conveniences to people’s daily lives,but also brought many security risks.At present,the security detection technology for Io T devices fails to adapt to its development speed,coupled with the shortage of its own system resources such as computing and storage that can be called,so there are many insecure devices in Io T that are vulnerable to attacks,which poses a serious threat to Io T security.The most well-known is the proliferation of Io T botnet.On the one hand,there has been a complete processing scheme in the detection field of traditional botnet,but the existing detection methods are not fully applicable because Io T devices are limited by their own computing,storage and other system resources.On the other hand,the rapid development of Io T has led to the emergence of various Io T botnet and their variants,while most existing methods ignore the impact of Io T botnet changes on detection models.Over time,this will inevitably lead to the existing detection model no longer applicable to the changing botnet.Based on the shortcomings of the above two aspects,this article has made improvements and innovations in feature dimensionality reduction and detection algorithms:1)An improved Fisher Score feature selection method was proposed,which filters out the optimal features based on the score and eliminates most irrelevant features.The experimental results show that this method performs well on four Io T datasets and significantly reduces the training time of the model.At the same time,experiments have shown that models with fewer optimal features can still achieve high accuracy,further reducing the system resources required for model training,making the detection model suitable for resource scarce Io T devices.2)A concept drift resistant algorithm based on gradient boosting decision tree improvement,GBDT-IL(Gradient Boosting Decision Tree incremental learning),is proposed to adapt to newly emerging data samples in data streams through incremental learning.At the same time,considering the overfitting caused by the redundancy of the incremental learning process,the process of pruning the tree is added to improve the model performance.The final experiment shows that compared to traditional machine learning algorithms,this method can improve the model accuracy by up to 34.8%.Compared to existing anti concept drift algorithms,this method also performs better.