Research On Detection Methods For Malicious Social Bot Based On Deep Neural Network

Online social networks have brought great convenience for people,at the same time their broad social and economic value has become the target of attackers to concoct a plot.Not only have malicious social bots in the online social networks polluted the social network environment,but also led public opinions and posed a threat to national security.In this thesis,we study the detection methods for malicious social bots for the reasons that find the malicious social bot accounts quickly,prevent attackers from using social bots to incite public opinions,threaten the stability of the society and the nation,and finally protect public safety effectively.First of all,this thesis summarizes and analyzes threats that malicious social bots in online social network have brought to personal information security,social public security and national security.Starting with the abnormal accounts in the early social network,we explain how these accounts gradually evolve into malicious social bots and have stronger destructive power.At the same time,we expound shortcomings of the existing detection methods for social bots in detecting feature selection,data annotation,models for detection and so on.Secondly,as existing researches are weak in the features for detection and give little attention to emotional features,we propose a systematic feature extracting system for the detection of malicious social bots.The features which contain account features,content features,emotional features,propagation features and evolutional features are divided into static features,dynamic propagation features and relational evolution features.We also propose a sentiment classification model based on LSTM network and Attention mechanism.The model takes advantage of LSTM network and Attention mechanism in extracting features to emphasize extraction of sentiment features for online social network accounts.Afterwards,most of the existing datasets are not labeled,and the manual annotation has great subjectivity,which affects final detection results.As a result,we propose an unsupervised method for dataset annotation based on parallel computing and implement it on the Hadoop platform.Not only can the method get rid of interference caused by subjective factors in the process of manual marking,but also can reduce cost of computation,and provide a more reliable dataset for subsequent detection.Finally,as the models of existing detection methods are too single and have poor performance for detetction,we propose a method for malicious social bot detection based on CNN,and give some measures for optimization from experimental datasets and model hyperparameters to further optimize the model.The experimental results show the feasibility and effectiveness of the model for malicious social bot detection based on CNN in this thesis.