Design of a Deep Learning Model for Cyberbullying and Cyberstalking Attack Mitigation via Online Social Media Analysis

Abstract

Identification of cyberbullying and cyberstalking for real-time use cases is a multi domain task that involves the design of social media data extraction, sentiment analysis, sentiment pattern evaluation, and regression models. To perform this task, researchers have proposed the use of high-density feature representation models that can extract social media sentiments, and combine them with user specific parameters like age, gender, time of post, etc. But existing models are either non-comprehensive or capable of achieving limited accuracy when used for real-time scenarios. Moreover, these models are not flexible to multimodal inputs, which further limits their scalability levels. To address these concerns, this paper proposes the development of a deep learning model for cyberbullying and cyberstalking attack mitigation via social media analysis. The proposed model initially collects tweets posted by users, extracts meta data, and analyzes language features for training a Long-Short-Term Memory (LSTM) based Convolutional Neural Network (CNN), which assists in the pre-filtering of tweets. The filtered tweets are passed through a Natural Language Processing (NLP) engine that assists in sentiment identification for these texts. Sentiment data and Word Embedding capabilities are used to anticipate cyberbullying and cyberstalking attacks. This is done via CNN based pattern analysis, which assists in the efficient identification and mitigation of these attacks. Due to the integration of these models, the proposed method is able to improve attack detection accuracy by 3.5 %, while reducing the identification delay by 8.3 % in real-time scenarios.