Comment by godelski

Actually I'm really curious how you do this. If the photos aren't stored client side, then how do you search? Do you have a thumbnail of every photo client side? Is that enough? I mean ImageNet scores are still pretty low for small/fast neural nets. And ImageNet isn't even representative of real world photos. So obviously to be successful you're going to have to continue training. So how do you do this in a privacy preserving way? Even federated learning can have some issues because images can be reconstructed from gradients.