Multinetguard: A Deep Learning Framework For Real-Time Multimodal Cyberbullying Detection In Social Media Using Bilstm And Cnn

Abstract

The growing incidence of cyberbullying on social media channels seriously endangers the well-being of users, particularly young adults and teens. While conventional detection techniques have concentrated on text-based abuse, memes and annotated screenshots—image-centric forms of bullying—are on the rise. We present MultiNetGuard, a deep learning-based multimodal system able to detect cases of cyberbullying by examining the visual and textual elements of social media posts, therefore tackling this difficulty. While the other half processes image data using a pre-trained Convolutional Neural Network (CNN), specifically ResNet-50, one component of the system handles textual input using a Bidirectional Long Short-Term Memory (BiLSTM) network with attention mechanisms. A multimodal fusion layer can more accurately classify cyberbullying material by means of cross-modal semantic links and context capture by combining characteristics from both branches. Our model was both trained and validated on a custom-labelled dataset that includes real-world social media memes, abusive comments, and image captions. Content flagging of the system with early-warning alerts creates proactive moderation and user protection. This study shows that visual and textual media give one fuller knowledge of negative purpose—thus promoting healthier, safer online communities.