Modern Vision-Language Model (VLM) has shown tremendous performance in multi- modal reasoning, captioning, retrieval and generative tasks, but it is also fatally suscep- tible to two types of attacks: adversarial ...