Research On Temporal Convolutional Network For Semi-supervised Internet Of Things Communication Anomaly Detection Model

In recent years,Internet of Things(IoT)has developed rapidly and a lot of intelligent terminal devices are deploied in Internet of Things communication network.However,there are some security problems,which are caused by device failure or attacks from the outside world.Because of frequent communication of devices,IoT may easily accumulate many communication data,which cannot be completely labeled.These data contain the communication and control information with some potential threats and anomaly.Moreover,there are complex anomaly type and strong time sequence relationship between communication data,which rises the difficulty of anomaly detection model.Meanwhile,Temporal Convolutional Network(TCN)achieves better performance for series problems compare to other neural networks.Hence,this thesis aims to propose semi-supervised model for solving IoT communication anomaly detection based on TCN,the main research contents are as follows:(1)This thesis proposes a semi-supervised anomaly detection model named Hierarchical Stacking Temporal Convolutional Network(HS-TCN).HS-TCN considers the temporal relationships of IoT communication data to utilize a small number of labeled data and a lot of labeled data.The application of hierarchical and stacking methods can train model step by step and remove the uncertain records during the training process.The experimental results show that HS-TCN achieve obvious improvement than supervised models and have better performance than other semi-supervised models.Morever,HS-TCN achieves the balance of model performance and complexity.(2)This thesis proposes Semi-Supervised Variational Temporal Convolutional Network(SS-VTCN),which is used to solve the multi-anomaly detection problem in IoT communication.SS-VTCN combines Variational Autoencoder(VAE)and TCN to learn the feature representation and time sequence relationship between communication data.SS-VTCN applicates TCN to predict the preliminary results.Meanwhile,the output of TCN is regarded as the VAE input for reconstructing data.Moreover,SS-VTCN calculates the reconstruction error according to the reconstructing data and VAE input to judge the preliminary prediction by rectification strategy and ensure the finally type.The experimental results show that SS-VTCN is more suitable for IoT communication anomaly detection problem than contrastive supervised models and has obvious performance advantages compared to other semi-supervised learning models.