neuralhash-collisions

Reminder that it's perceptual hashes, not cryptographic hashes. So it's enough that images have enough visual resemblance, according to the model. Natural collisions have been observed and generating collisions for planting on someone elses device is trivial. Apple also only recieves a DB of hashes and so have no way to verify that they're only scanning for CSAM and not other "undesirable" content. https://github.com/roboflow-ai/neuralhash-collisions https://github.com/anishathalye/neural-hash-collider

The sample set is ImageNet, which is a well-known dataset in Computer Vision and is available for download here: https://www.kaggle.com/c/imagenet-object-localization-challe... I'd love to see this work extended; if you find additional collisions in the wild please submit a PR to the repo (please do not submit artificially generated adversarial images): https://github.com/roboflow-ai/neuralhash-collisions For what it's worth, Apple claimed to find a _lower_ incidence of false-positives when it used pornographic images in its test[1] (which makes sense; images containing humans is probably more aligned with what the model was trained on than nematodes) [1] https://tidbits.com/2021/08/13/new-csam-detection-details-em... > In Apple’s tests against 100 million non-CSAM images, it encountered 3 false positives when compared against NCMEC’s dat