Research On Social Spammer Detection Based On Modified DCNN

With the development of the Internet,online social networks have already become very important part of our life.Its appearance not only lets our communication more and more convenient,but also makes our life more and more rich.However,online social networks also provide a platform for social spammers to spread malicious information and links.According to an effective study in 2013,about that one out of every two hundred social messages is malicious information or links spread by social spammers.Through the research,we found that most of the existing detection methods to identify social spammers mainly learn the classification model by analyzing and extracting the user's behavior information.However,due to information security and personal privacy issues,online social networking sites can't monitor all behavioral of users.Therefore,the behavior information we get from social networks is not comprehensive enough.In addition,we also found that social spammers are constantly changing and upgrading themselves in order to better avoid the detection of the system.They disguise themselves as legitimate users by using a diversity and variability of social spam sending strategies in an attempt to avoid detection.This makes it difficult for existing detection methods to distinguish social spammer from legitimate users only by using their own behavior information.In order to solve this challenge,in this paper,we propose to use a novel,graph-based classification model,namely the diffusion convolutional neural network model,to detect social spammers.Different from other detection methods,diffusion convolutional neural network model can learn behavior information from other users through graph structure(i.e.social network relationship)and use it to identify social spammers.However,the original diffusion convolutional neural network model is a general classification model.In order to make the original diffusion convolutional neural network model more effective for detecting spammers,we modify it by using information attenuation Coefficient and social Regularization,which is called DCNN+CR model.Finally,the experimental results obtained by using the real datasets obtained from Twitter show that our DCNN+CR model is not only superior to the original diffusion convolutional neural network model,but also outperforms the existing spammer detection methods,especially in terms of accuracy and F1-score.