@xfinity

Adversarial examples—images with subtle, human-imperceptible noise—that trick a vision language model (Llava 7B) into misidentifying the Xfinity logo as a hate symbol. The process uses gradient ascent…