Multimodal Cyberbullying Detection on Social Media Using LLMs

This study explores automated cyberbullying detection across major social networks and messaging platforms using state-of-the-art large language models in a zero-shot, multimodal setting. Models including LLaMA 4, Gemma 3, and GeminiAI were evaluated on images and videos without domain-specific fine-tuning. The system assigned a continuous score (0–10) to indicate the presence of cyberbullying across four categories: revenge porn, happy slapping, racism, and body shaming. Experiments on over 5,000 multimedia samples from Telegram, Reddit, and X (formerly Twitter) showed that large language model-based approaches achieve competitive performance, with Gemma 3-12B emerging as the most stable, accurate, and ethically compliant model. The results also highlighted the critical role of prompt engineering and multimodal context in detecting subtle or implicit online aggression.