If Apple is indeed using CNNs, then I don’t see why any of the black-box adversarial attacks used today in ML wouldn’t work. It seems way easier than attacking file hashes, since there are many images in the image space that are viable (e.g., sending a photo of random noise to troll with such an attack seems passable).