X-MCB-Net: Explainable multimodal Cyberbullying Detection Network using Swin Transformer

Cyberbullying on social media represents a persistent digital safety concern, often conveyed through the interplay of offensive text and images. Existing detection systems largely treat visual and textual modalities independently and offer limited interpretability, hindering deployment in accountable content-moderation workflows. This study presents X-MCB-Net, an explainable multimodal cyberbullying detection framework that integrates the Swin Transformer for visual feature extraction with XLM-RoBERTa and a bidirectional GRU for textual analysis. A dynamic intermodal attention mechanism fuses visual and linguistic representations to capture implicit and context-dependent cues for abuse. The core novelty lies in the joint design of adaptive cross-modal fusion and explainability-aware attribution, enabling both high detection performance and transparent decision support for human moderators. Evaluated on 5790 annotated image–text pairs from the Multimodal Cyberbullying Detection Dataset, X-MCB-Net achieved 98.8% accuracy and 96.6% F1-score under stratified 5-fold cross-validation. Explainable AI (XAI) visualizations identify the image regions and text tokens that drive each classification decision, supporting auditability and moderator trust in automated online safety systems. Future work will extend the evaluation to additional multimodal benchmarks and multilingual corpora to establish the framework’s generalizability across diverse platforms and cultural contexts.