New GeoThreat attack targets Large Vision-Language Models in remote sensing

Researchers have developed GeoThreat, a novel adversarial attack method designed to manipulate Large Vision-Language Models (LVLMs) in the context of remote sensing image interpretation. This method targets the models' ability to understand both local details and global scene context, which is crucial for analyzing remote sensing data. GeoThreat operates by modulating adversarial representations at both conceptual and perceptual levels, using class tokens from surrogate image encoders and patch tokens from adversarial examples. Extensive experiments have shown GeoThreat to be superior in both transferability and controllability when attacking various LVLMs. AI

IMPACT This research highlights potential vulnerabilities in LVLMs, particularly for specialized applications like remote sensing, and could spur development of more robust models.

RANK_REASON Research paper detailing a new adversarial attack method. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CV →

• arXiv
• GeoThreat
• Large Vision-Language Models
• remote sensing

• paper
• safety

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