Exposing Vulnerabilities in Visible-Infrared VLMs: A Unified Geometric Adversarial Framework with Cross-Task Transferability
Researchers have developed CFGPatch, a novel adversarial framework designed to expose vulnerabilities in visible-infrared vision-language models (VLMs). This method utilizes curved-edge fractal geometry and a modality-specific rendering mechanism to create adversarial patches that disrupt both shape and texture perception in VLMs. Experiments demonstrate that CFGPatch effectively fools these models and shows strong transferability across different tasks like image captioning and visual question answering. AI
IMPACT This research highlights potential security risks in multimodal AI systems operating in challenging environments, suggesting a need for more robust adversarial defenses.