Research On Detection Methods For Cyberbullying Based On Deep Neural Network

Cyberbullying refers to the behavior of posting unfavorable information to others through digital media in the Internet environment with the purpose of hurting or embarrassing others,thereby causing mental and psychological harm to others.With the rise and popularity of social networks,cyberbullying incidents occur frequently.Cyberbullying has become the primary threat to teenagers' online security,and has attracted the attention of governments and researchers all over the world.This thesis studies the automatic detection technology of cyberbullying in social networking sites,aiming to detect cyberbullying accurately and effectively.At the same time,the root cause of the negative impact of bullying on social networks is curbed,and the online safety of social network users,especially young users,is guaranteed.Firstly,this thesis analyzes the causes and effects of the worsening cyberbullying based on the current development of the Internet and social networks,and comprehensively summarizes the threats cyberbullying brings to the online security of social network users,especially teenagers.This thesis analyzes several existing works in the field of cyberbullying detection based on machine learning or deep learning,and summarizes the shortcomings of these existing works in detection feature,model structure,privacy protection,etc.Subsequently,when it comes to the excessive reliance on textual features in the existing bullying feature selection work,textual,emotional and user-level features are proposed to be integrated as the final bullying detection features.The textual features are obtained based on the vicious word dictionary and topic keywords,the emotional features are extracted by an ELMO-based emotional feature extraction model,and the user-level features are supplemented by means such as crawlers and API.Afterwards,different from the existing bullying detection model which generally has a single structure,RNNs-based model is constructed in feature extraction.At the same time CNN-based model is constructed as the classifier in the task of cyberbullying detection.When it comes to the results of classifier on different datasets,balancing data and Dropout are adopted as optimization strategies.The feasibility and effectiveness of the cyberbullying detection model based on deep neural network proposed in this thesis are verified through experiments.Finally,federated learning is introduced to ensure the privacy security of data.According to the general application framework of federated learning and its application process in CNNs-based model,a cyberbullying detection model based on federated learning is proposed.Then a specific implementation scheme based on the Pytorch is given,which ensures the privacy security of data in the process of training under the premise of ensuring certain model accuracy.