Abstract: In this paper, we present a novel task of source-free cross-modal adversarial example generation, which generates adversarial examples based on textual descriptions of attackers. This task ...