New QR-Structured Thermal Triggers Attack Infrared Vision-Language Models

Researchers have developed a novel framework called QR-Structured Thermal Triggers (QR-STT) designed to manipulate infrared vision-language models (IR-VLMs). This training-free, black-box method uses QR patterns with varying thermal states to steer the model's semantic understanding. Experiments demonstrate that QR-STT can successfully redirect IR-VLMs towards specific concepts, impacting tasks like image captioning and visual question answering while maintaining visual stealth. AI

IMPACT Highlights a new attack vector for vision-language models, emphasizing the need for improved robustness evaluations.

RANK_REASON Academic paper detailing a new method for attacking AI models. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• arXiv
• CLIP-style encoders
• Hugging Face
• Infrared vision-language models
• IR-VLMs
• QR-Structured Thermal Triggers
• QR-STT

• paper
• safety

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →