Methodology for Adversarial Attack Generation: Using Directives to Mislead Vision-LLMs

This article details the multi-step typographic attack pipeline, including Attack Auto-Generation and Attack Augmentation.