This research introduces a novel framework that combines knowledge distillation from Large Visual Language Models with knowledge infusion from ConceptNet to improve toxicity detection in multimodal content like memes.

• Achieves significant performance gains over baseline models by leveraging both LVLM capabilities and external knowledge graphs
• Creates a more robust detection system that better understands contextual connections across text and images
• Demonstrates how knowledge models can be effectively combined to address complex security challenges
• Provides a practical approach to creating safer online environments through advanced AI techniques

Why it matters: As online toxicity evolves to use subtle combinations of images and text, these advanced detection methods become crucial for content moderation systems and protecting users from harmful content.

Just KIDDIN: Knowledge Infusion and Distillation for Detection of INdecent Memes