Research On The Key Technology Of Spammer Detection In Social Networks

The progress of mobile Internet technology promotes the vigorous development of social networks,meanwhile social networks have been active places for spammers.Without the permission of the recipient,spammers send information to normal users in large quantities,seriously threatening the information security of normal users and the credit system of social networks.As a consequence,the spammer detection method in social networks has become a research hotspot,which has significant applied value in widespread fields such as e-commerce recommendation,information retrieval,and network security.However,there still exist some shortcomings and limitations in the spammer detection,which has been widely applied at present:(1)Although the behavior of individuals is complex and random,the characteristics of user behavior on the group level shows periodicity to a certain extent,which wasn't considered by the current algorithms of detecting spammer active time based on time,consequently,causing the normal time to be mistaken as active time.(2)Most of the spammer detection methods based on network topology only focus on using local location information,lacking of in-depth mining of global location information.(3)The current spam detection algorithms based on text generally face problems of many noisy words existing extensively in the text,and too sparse semantic features extracted from the text.In view of the above issues,a two-stage detection process has been proposed.At first,considering the concentrated active time,we utilize time characteristics to detect active time of spammers.In the detected active time,the spammers are detected through topological and text features according to abnormal topological structure and irrelevant text content.The main contents and innovations of the research are listed as follows:1.A spammer active time detection algorithm based on Poisson process is proposed.Primarily,we cut the data set to a certain time granularity,and calculate the sequence of multigranularity network structure indicators.Secondly,the classical Poisson process is improved in three aspects: periodic integration,continuity expansion and comprehensive detection,forming the spammer active time detection algorithm.The experimental results show that the algorithm has better detection effect and a more effective combination of network structure indicators,which is helpful for further detection work in the future.2.A spammer detection algorithm based on graph convolution network is proposed.At first,we utilize network representation learning methods to extract the local location information of users in the network.What's more,the graph convolution network is proposed based on orthogonal polynomials which can approximate the convolution operation in the spectral domain,to deeply mine the global location information and further detect spammers.Compared with other methods,the proposed method preforms higher accuracy and faster efficiency,which is suitable for largescale data sets and data sets with sparse labels.3.A spam detection algorithm based on attention mechanism is proposed.First of all,on the basis of the original CNN model,a filter layer is added in which an attention mechanism based on Naive Bayesian weighting technology is designed through integrating the idea of keyword extraction.By reducing noise words in the text,the noisy issue is solved and the candidate set of words with detection effect is screened out.Moreover,a pooling strategy based on attention mechanism is proposed,giving high weight to words with detection effect,extracting the effective feature representation and alleviating the sparse representation problem.According to the experiments,the algorithm shows high accuracy and robustness.