Detection And Prevention Of Malicious Accounts And Messages In Online Social Networks

Social networks become an important part of people’s life.While people enjoy the convenience brought by social networks,they also face the security threat from social network.Through social networks,attackers can spread malicious URLs more quickly and widely.In order to provide users with a secure social network environment,researchers propose solutions mainly from three directions,including the detection of malicious accounts,the detection of malicious messages and the prevention social network attacks.Existing works mainly focus on detection of malicious accounts and messages.The main detection technologies include detection algorithms based on social graph and detection methods based on machine learning.Most detection algorithms based on social graph are computationally intensive and can only run offline.Therefore,it is necessary to design some lightweight algorithms.The detection method based on machine learning needs to find the features that are difficult for attackers to change,otherwise once the attacker adjusts the attack strategy,the detection method will be invalid.In addition,the prevention of social network attacks has gradually attracted the attention of researchers.Since the behavior pattern of normal users is relatively stable,the risk assessment of normal users can be used to prevent attacks before they are launched.The paper starts from the following three directions to detect and prevent social network attacks.Aiming at the detection of malicious accounts,we establish a graph structure based on the relationship between users and friends,and designs a detection algorithm to distinguish normal users from malicious accounts by analyzing the location characteristics of users in the graph.In order to spread malicious messages as widely as possible,attackers adopt different friend policies from normal users to establish friend relationships with a large number of users.In this paper,the detection algorithm is evaluated on the Sina Weibo dataset,and a threshold value is calculated to distinguish normal users from malicious accounts.Experimental results show that the accuracy of the detection algorithm in this paper is 86.27%,and the false alarm rate is 8.54%.Aiming at the detection of malicious messages,we establishe a text classification model based on deep learning to classify normal messages and malicious messages.The accuracy of the model is 91.36% and the false alarm rate is 8.82%.In addition,in order to improve the detection efficiency as much as possible and realize real-time detection,this paper proposes a multistage detection framework.By deploying a lightweight detection model on the edge of the network to pre-tag the text information sent by the user,the server can give priority to detecting the labeled suspicious samples when the server’s computing resources are limited,so as to improve the utilization rate of the server’s computing resources.Aiming at the prevention of social network attacks,we proposes a user analysis model.First,three kinds of features are extracted from users’ social information: statistical features,social graph features and semantic features.These featuress are used as input to the user analysis model and a security risk value is calculated for each user.Through security training and focused monitoring of high-risk users,the risk of an attacker using social networks to infiltrate the company’s internal network can be reduced.Experimental results show that the user analysis model can effectively evaluate the security of users’ activities in social networks.This paper studies from the above three directions and proposes a new solution,which not only detects malicious messages and malicious accounts,but also assesses the risks of normal users and makes efforts to provide users with a safe social network environment.Through experiments,we prove that these works are effective and have certain reference significance for similar work.