Moderating Illicit Online Image Promotion for Unsafe User-Generated Content Games Using Large Vision-Language Models
Researchers have developed UGCG-Guard, a system designed to identify illicit image-based promotions of unsafe user-generated content games (UGCGs). This system utilizes large vision-language models (VLMs) with a novel conditional prompting strategy and chain-of-thought reasoning to adapt to the unique nature of these images. UGCG-Guard achieved a 94% accuracy rate in detecting such promotional images in real-world scenarios, addressing a critical gap in online safety for children and adolescents. AI
IMPACT This system offers a new method for moderating harmful content in user-generated games, potentially improving online safety for young users.