A new framework named HMGUARD uses adaptive prompting and chain-of-thought reasoning within multimodal large language models to detect harmful memes.

The research is presented at NDSS 2025, with online resources listing the authors, presenter details, and a blog post that summarizes the work.

The study analyzes visual arts and propaganda techniques to explain why current detection methods fail and to guide the design of HMGUARD.

The article highlights how harmful memes spread on social media, emphasizing challenges in detection due to varied expressions, complex compositions, propaganda strategies, and cultural contexts.

HMGUARD reportedly delivers strong performance, achieving 0.92 accuracy on a public harmful meme dataset and outperforming baselines by 15% to 79.17%, while surpassing existing tools with 0.88 accuracy in real-world use.