Research On Passive Flow Detection Method In Internet Governance

With the development of communication technology and the times' progress,the amount of network access to mobile phones,computers,TV and other multimedia devices has increased.All kinds of network activities produce text,audio,video,file transmission and other traffic data.Besides,many applications have different requirements for the quality and quantity of traffic data in Internet.The rapid growth of traffic has brought considerable challenges to Internet governance,requiring research in multiple directions such as active detection technology and passive traffic detection.The research content of this paper belongs to the research category of passive flow detection methods,which can distinguish the application of the flow.Accurate traffic classification is the primary task of Internet governance.Most of the traditional traffic classification methods are port-based and payload based.These methods rely on the explicit characteristics of traffic and have weak classification ability for encrypted traffic.Traffic classification algorithm based on statistical features has made significant progress,but it needs to select features manually,which is not conducive to dealing with complex classification task scenarios.In this paper,the deep learning method that can automatically extract the classification features is used to build the traffic classification model.The deep reinforcement learning module is introduced to optimize the distribution of flow sequence features,which has better classification performance in large traffic scenes.The main work of the article includes:The two-stage encrypted traffic classification model proposed in this paper can complete the classification training using the packet header data in the original traffic format and a small amount of encrypted payload data.The model is constructed according to the three-level structure of traffic.In the first stage,the convolutional network is used to learn the spatial structure of data packets and generate the feature vector of data packets.In the second stage,the convolution network is used to learn the temporal form of data packets and generate the flow feature vector.The experimental results show that the training set's classification accuracy is 97%when the proposed model is used in the multi-application classification task of the VPN-non VPN dataset.Compared with the baseline method,the proposed model needs fewer sample data.The RD-combine model proposed in this paper has the self-learning characteristics and extract classification features automatically.For the large traffic classification scenario,the deep reinforcement learning module can filter the data packets in the preprocessing stage and retain the data packets with more classification features.The deep reinforcement learning module can optimize the feature distribution and classify large traffic series more accurately.Firstly,the convolution neural network is used to extract the packet features,and the strategy gradient method of reinforcement learning is used to filter the stream sequence.Finally,the GRU network is used to train the classification model of the filtered stream sequence.Experimental results show that the RD-combine model can significantly shorten the traffic sequence's length and has similar classification accuracy.The model has a faster training speed and convergence rate.