Multimodal Cyberbullying Meme Detection From Social Media Using Deep Learning Approach

Cyberbullying includes the repeated and intentional use of digital technology to target another person with threats, harassment, or public humiliation. One of the techniques of Cyberbullying is sharing bullying memes on social media, which has increased enormously in recent years. Memes are images and texts overlapped and sometimes together they present concepts that become dubious if one of them is absent. Here, we propose a unified deep neural model for detecting bullying memes on social media. In the proposed unified architecture, VGG16-BiLSTM, consists of a VGG16 convolutional neural network for predicting the visual bullying content and a BiLSTM with one-dimensional convolution for predicting the textual bullying content. The meme is discretized by extracting the text from the image using OCR. The perceptron-based feature-level strategy for multimodal learning is used to dynamically combine the features of discrete modalities and output the final category as bullying or nonbullying type. We also create a “bullying Bengali memes dataset” for experimental evaluation. Our proposed model attained an accuracy of 87% with an F1 score of 88%. The proposed model demonstrates the capability to detect instances of Cyberbullying involving Bengali memes on various social media platforms. Consequently, it can be utilized to implement effective filtering mechanisms aimed at mitigating the prevalence of Cyberbullying.