Malicious Account Classification In Online Social Network Based On Deep Learning

Malicious detection is the problem of predicting abnormal accounts or nodes in online social networks(OSN).Since this problem applies to various tasks(such as malicious URLs or user content classification),it has attracted widespread attention from researchers in the computer security field and the problem of identifying malicious accounts has also been widely studied.However,conventional techniques such as rule-based,statistical methods,and even general learning methods cannot deal with the dynamic environment with increasing users and data volumes.Therefore,a novel learning model for detecting malicious content types(such as spam and malware)based on user characteristics analysis has become one of the current research areas.This thesis focuses on the classification of malicious accounts,mainly using binary classification concepts and learning methods.First,this study collected a wide range of OSN feature data,including user configuration files,text content,and URL information as an experimental dataset.On this basis,using deep learning techniques that have achieved widespread success in many fields,this paper studies a learning classifier with generic functions to optimize the classification results.At the same time,this article conducts an in-depth study of many aspects of the malicious account problem.Finally,this paper's method is compared with some current mainstream methods,which shows that our method has higher accuracy in identifying malicious accounts.The main contributions of this article in the classification of malicious accounts include the following aspects.Firstly,a dynamic CNN model with a large-scale OSN data set is proposed,including a URL list,user comments,and user profile to build a benchmark dataset.Using this dataset,we construct an effective classification model by extracting various OSN features.To identify malicious account problems,a supervised learning method is used to distinguish between malicious users and normal users.Secondly,this study proposes a deep learning classification model with a pooling function.Unlike the general classifier model,this paper designs a dynamic pooling function in the hidden layer of the deep learning network model to improve the accuracy of neural network training.URL classification is a crucial element for malicious account detection before the pre-connection phase in the OSN.To this end,the method is applied to URL feature datasets to classify URLs.According to the experimental results,the deep learning classification model of RunPool pooling operation with Gaussian function can produce higher accuracy and small loss scoreThirdly,we present a learning method for classifying malicious comments with a generic regularization function called RunOut.Applying the regularization function to the deep learning classification model's hidden layer can effectively solve the overfitting problem of the neural network and improve the training result of the model.Experiments show that the model can achieve good results in detecting malicious accounts using the user comment dataset.Finally,this study develops a non-linear activation function for the classification model called RunMax.This part of the thesis develops a function named RunMax as the activation function in the last layer of the hidden layer.Based on the experiment result after training the user attribute datasets,the activation algorithm can improve neural networks' performance in the classification of fake profiles issues.Compared with other learning algorithms under the same adjustment parameters and experimental environment,the RunMax can produce better performance by obtaining higher detection accuracy and tiny loss for fake accounts detection.