Analysis Of Audio And Video Encrypted Traffic Based On Machine Learning

With the rapid development of mobile Internet,the growth of short video applications is particularly rapid.Reasonable traffic configuration management is an important part.Audio and video traffic identification has become a research hotspot in academia and industry in recent years.Among them,SSL/TLS encrypted audio and video applications are becoming more and more complex,and accurate traffic identification and classification are vital to effective network management and resource utilization.Based on the analysis of traffic characteristics,machine learning and deep learning are used to improve and innovate the application recognition research in the audio and video field.Aiming at the feature analysis of audio and video traffic,a semi-supervised feature selection algorithm(SRR-LSA)is proposed based on improved Relief F and ACO algorithms.The improved Relief F algorithm improves the limitation that the original Relief F algorithm only evaluates the importance of a single feature,can quickly reduce the dimensionality of the feature and provide good prior knowledge for the improved ACO algorithm.In the feature extraction,the candidate features are randomly combined after the Relief F algorithm for weight calculation and Spearman correlation coefficient is used to perform correlation analysis to initially eliminate redundant features.The improved ACO algorithm searches for the optimal feature subset by continuously searching the feature space.Compared with the original ACO algorithm,LS feature value is used on pheromone update step after each iteration for secondary correlation analysis;Aiming at the identification of SSL/TLS encrypted audio and video applications,a framework for the fusion of encrypted traffic and non-encrypted traffic is proposed.For non-encrypted traffic,Bayesian classification is adopted based on flow sets and the flow set data generated by correlation and classification of flow data is used as the input data set.For encrypted traffic,a model input for the combination of feature engineering and fixed-length timing packets for deep learning to identify encrypted traffic is proposed.The feature vector of a single data packet is added in addition to some of the characteristics of the identification flow and the header information of the data packet.In the first 60 bytes of application layer data,the packet vector group is a sequence composed of 5 packet characteristics.Then,CNN is used to learn spatial features and LSTM is used to learn timing features.The model weights the output probabilities of the non-encrypted module and the encrypted module to obtain the final probability estimate.Experiments show that the feature selection algorithm can effectively reduce the complexity of feature subsets and achieve a better classification effect,and the fusion identification framework has better classification performance and generalization ability.