Multi-Model Fusion for Enhanced Cyber-Bullying Detection in Digital Communication

The advanced multi-modal approach is crucial in cyberbullying detection since it allows a complex examination of three elements: texts, images, and videos. It uncovers fine textures of bullying that the mono-modal system cannot find. It provides precision in detection and assurance of its effectiveness in the range and development of cyberbullying forms in communications. This research introduces an improved multi-modal framework (IMMF) that deals with text-based, visual and video content. An incorporated system uses cutting-edge technologies like vision transformers (ViTs), where a self-attention mechanism is applied comprehensively to get a clear perception of spatial hierarchies, and graph neural networks (GNNs) for video content analysis, allowing them to track dynamic behaviors in the way video sequences. For textual content, the framework builds upon recent breakthroughs in natural language processing with BERT (Bidirectional Encoder Representations from Transformers), allowing the approach to capture discourse-level semantic cues, and subtly revealing aggression. These holistic measures, in turn, improve feature extraction in different media types and enable capturing the most complex web of harassment with increased accuracy. Furthermore, the system uses a multi-modal fusion method, combining the information gathered from various modalities, thus providing an enhanced perception of the situation and improving detection accuracy. The model develops few-shot learning (FSL) techniques that help the model adapt to new types of bullying with the help of minimal additional data to keep pace with the rising threats in new forms of bullying. The amplification of the system reached a 98% accuracy generalization and up to 7.5%, including a noticeable diminishing of the false alarms. Such results point out the necessity of smart computational models like ViTs, GNNs, and FSL for the fight against cyberbullying, which is the primary thing on the way to the successful development of a safe and caring online social environment.