New adversarial patch UCGP exploits infrared vision-language model vulnerabilities

Researchers have developed a new adversarial patch framework called Universal Curved-Grid Patch (UCGP) designed to exploit vulnerabilities in infrared vision-language models (IR-VLMs). Unlike previous methods, UCGP targets open-ended IR-VLM tasks, aiming to degrade classification, captioning, and visual question answering simultaneously with a single deployable artifact. The framework uses a low-frequency curved grid and optimizes it through a representation-driven objective that disrupts the visual representation space. Experiments demonstrate that UCGP can effectively degrade IR-VLM performance across various architectures and generalize to different datasets and real-world scenarios, highlighting a significant robustness issue in current infrared multimodal systems. AI

IMPACT Highlights a critical vulnerability in infrared multimodal systems, potentially impacting their reliability in low-visibility applications.

RANK_REASON Academic paper detailing a new adversarial attack method for IR-VLMs. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CV →

• Expectation over Transformation
• infrared vision-language models
• IR-VLMs
• Meta Differential Evolution
• Thin plate spline
• UCGP
• Universal Curved-Grid Patch
• Yuxian Dong

• paper
• safety

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