Research On Intelligent Analysis And Profiling Methods Of Internet End Target

An Internet end target is a device or a user located at the edge of Internet,and the profile of network end target is a collection of attributes which can fully depict the end target.Intelligent analysis of the Internet end target is an effective method to construct end target profiles.Such a behavioral profile is an essential information basis for almost any analytical applications of the end target.Thus,researches on the behavioral profile of the network end target have attracted a broad attention from both the academia,the industry and the society,and now it is part of the research frontier in fields such as network management,network security,electronic commerce and so on.For the research on the profile of Internet end target,multiple challenges are present.The most pressing challenge lies in the traffic-behavioral profiling of the Internet end target,where the privacy-infringement issues have occurred since most of the existing methods inspect the payload of the traffic data.Though,there are light-weighted analysis which leverages only the header information,the classification targets involved in these researches are overly simple and unitary.For the information utilization,there is still a wide range of potential information for mining,which is currently remain hidden.At the mean time,over-reliance on the manually-designed behavioral features is killing the gain in accuracy.Considering the application of behavioral profile,the non-stopping complication of application scenarios adds growing burden on the performance of the analytical methods,and the gap between state-of-the-art techniques and the application scenarios need to be bridged,so that the techniques can be implemented.To address the aforementioned challenges,this dissertation presents a deep dive into the research on Intelligent Analysis and Profiling of Internet End Target.The detailed contributions of this research are listed as follows:1.For the traffic-behavioral profiling of the end target,this research proposes the traffic classification based on higher-order statistical analysis,so that the limitation on traffic information mining can be broken through.This method introduces the higher-order statistical analysis into the extraction of network flow feature,where the formerly hidden information of the non-Gaussianity in network flow can be utilized.At the same time,in order to counter the non-stationarity of flow sequence,time-frequencyanalysis conceptions is adopted for better modeling to network flow.By combining the higher-order statistical analysis and the time-frequency conception,effective features are extracted from the traffic data.Together with the empirically optimized classifier,accurate classification of network traffic is achieved.For the building of the behavioral profile,this research proposes a series of traffic-behavioral features,which enriches current feature pool for use.At the mean time,this research provides the audience with a strategy for the learning of traffic-behavioral embedding representation,where together the end-target representation can be constructed in order to provide the following analysis tasks with refined information.2.For the recognition of the demographic attributes of end target,this research proposes the end-target demographics classification based on hierachical neural network.Neural network is a remedy for the over-reliance of hand-crafted features,and the hierachical structure of the network paves the road to a deeper inspection of the demographic information.Last but not least,integrated regularization method is proposed not only for an end-to-end training of the whole network,but it also grants controllability to the network structure during training.Experiments based on real-world dataset show that the proposed method out-performs several traditional methods on multiple metrics.3.For the recognition of the mental attributes,this research proposes the sentiment polarity classification based on attentional-graph neural network.This method is based on a hierachical neural network,which relieves the over-reliance on single category of information by fusing the content-text information and the user-connection information during the embedding phase,and thus an optimized usages of multiple information sources is achieved.Experiments based on real-world Twitter dataset exhibits the better performance for the proposed method than those for several traditional methods and a decision-stage fusion mechanism.4.For the recognition of the behavioral attributes,this research proposes a traffic-behavioral anomaly detection based on community discovery for end target.Rather than the well-researched detection of collective anomaly,this method focuses on the anomaly detection problem on single end target,which is as far as we know the first attempt made to such research field.This method utilizes a series of features to generate behavioral samples,and the behavioral graph can be built according to the distance of between samples.Community discovery method is adopted to find communities insidethe behavioral graph,which is followed by a threshold-based mechanism for anomaly detection.Apart from the aforementioned contributions,this dissertation also presents a number of experiments based on real-world dataset,where the empirically-optimized settings for the key parameters of the methods are given.The proposed methods are contrasted against multiple traditional methods,and the experimental results are analyzed.