Research On Detecting Spammers In Online Social Networks

The rapid development of social networks does satisfy human social requirements,but they are also soil for internet fraud.Using the characteristics of information propagation,spammers not only destroy someone’s reputation and privacy but also threaten social stability and national security.Spammers are good at mimicking the behaviors of legitimate users and always change their spamming strategies,which may make existing social spammer detection approaches failed.Towards this end,this thesis proposes using matrix factorization and deep learning techniques as well as considering the multiple types of information of social networks simultaneously to design effective models for detecting spammers:(1)Convex nonnegative matrix factorization(CNMF)-based social spammer detection.Most existing methods directly use handcrafted features to train classifiers.However,the extracted feature matrices are usually sparse,which makes their detection performance unsatisfactory.Moreover,how to use the social relations among users reasonably is still an open challenge of the social spammer detection task.To deal with these issues,we propose to learn user latent features and combine social interactions when training a classifier.In particular,we propose to CNMF to factorize legitimate user feature matrix and spammer feature matrix,respectively,which can avoid the issue of directly using handcrafted features.Besides,social spammers usually generate fake social relations to avoid being detected.To solve this issue,we propose a new social interaction-based regularization term in the loss function,which can further boost the detection performance.(2)Multiview-based social spammer detection.Current approaches only focus on a single type of information to detect social spammers.However,spammers usually use several types of strategies leading to the performance decrease of existing approaches.To tackle this problem,we propose a novel multiview-based social spammer detection model,which uses multiple types of information and social relations together to train a classifier iteratively.By factorizing different feature matrices simultaneously,the proposed model can automatically learn a consensus user matrix,which can effectively detect users utilizing multiple spamming strategies.We also propose a new approach to deal with view-level information incomplete issues,which further improves the detection performance.(3)Deep learning-based social spammer detection.Existing studies mainly use handcrafted features as the model inputs,making important partial information missing.Even for deep learning-based approaches,most of them only focus on detecting whether a single tweet is a spam or not,instead of spammer detection.To address these issues,we propose a novel attention-based social spammer detection model,which takes original tweets as the input.In particular,we propose a new attention mechanism to automatically distinguish the importance of each tweet when learning user content representations.Moreover,we design a new interaction frequency-based graph attention mechanism in the graph attention network.By considering both content-level and network-level features,the proposed model can improve the detection performance.(4)Deep learning-based implicit spammer detection.Propagating spam information using images is new type of spamming strategy,and such spammers cannot be detected by existing feature-based or content-based approaches.To solve this issue,we propose to use deep learning techniques working with frequency domain pre-processing techniques to detect such implicit spammers.Experimental results validate that spam information usually exists in the highfrequency domain regions.To address this problem,we propose to use Haar wavelet transform to filter the low-frequency domain regions in images,which can accelerate the convergence.Moreover,using prior knowledge to select particular modules and the appropriate number of layers can further enhance the performance of detecting implicit spammers.