1 articletagged with “vision-model”
Attack vision-language models through adversarial images with embedded text, perturbations, and visual trojans.