UGCG-Guard uses VLMs to detect illicit game ads with 94% accuracy
A new AI system flags sexually explicit and violent content in UGC game promotions targeting kids.
A team of researchers from the University at Buffalo, University of Notre Dame, and other institutions have introduced UGCG-Guard, a cutting-edge system designed to automatically identify illicit image promotions of unsafe user-generated content games (UGCGs) on social media platforms. UGCGs, popular among children and adolescents for their creative and social features, are increasingly being exploited by game creators who post sexually explicit and violent promotional images on platforms like Twitter and Instagram, inadvertently attracting young users. The researchers collected a real-world dataset of 2,924 such images, revealing a previously unstudied threat.
UGCG-Guard leverages recently introduced large vision-language models (VLMs) with a novel conditional prompting strategy for zero-shot domain adaptation, combined with chain-of-thought (CoT) reasoning for contextual identification. The system achieves an impressive 94% accuracy in detecting these illicit promotional images in real-world scenarios. The work, accepted at USENIX Security '24, represents the first systematic study of this online safety issue and provides a practical tool for social media platforms to protect minors from harmful content.
- Dataset of 2,924 real-world images featuring sexually explicit and violent UGC game promotions.
- Uses large vision-language models with conditional prompting and chain-of-thought reasoning for zero-shot detection.
- Achieves 94% accuracy in identifying illicit UGCG promotions in real-world scenarios.
Why It Matters
Automates moderation of hidden online ads targeting kids, addressing a gap in child safety tools.