Typographic Attack

Typographic attacks exploit the vulnerability of vision-language models (VLMs) to misleading text overlaid on images, causing misclassifications or incorrect reasoning. Current research focuses on understanding the susceptibility of various VLMs, including large language models and those used in autonomous driving, to these attacks, developing methods to generate more effective attacks (e.g., using reinforcement learning and self-generated attacks), and exploring defense mechanisms such as prefix learning. This research is crucial for improving the robustness and security of VLMs, particularly in high-stakes applications where misclassification could have serious consequences.