Hash collision in Apple NeuralHash model

Re:Hash collision in Apple NeuralHash model by avsteele on Wednesday August 18, 2021 @08:08AM (#28220057)

Apple claimed 'one in a trillion' chance of a collision. This is a great example of why you should not trust such an assessment.

[ Parent | on HN ]

Re:Hash collision in Apple NeuralHash model by manmal on Wednesday August 18, 2021 @09:17AM (#28220571)

That number is based around their plan to only trigger a search if a certain number of CSAM hashes were detected on a phone. A single collision wouldn't trigger such a search. A sufficiently motivated person can of course get access to a phone and put plenty hash-colliding images in the library.

[ Parent | on HN ]

Re:Hash collision in Apple NeuralHash model by cyanite on Wednesday August 18, 2021 @04:34PM (#28226739)

That’s a different scenario. That’s for it happening by chance, not on purpose. At any rate, the “matching” images are reviewed (or a “visual derivative” is).

[ Parent | on HN ]

Re:Hash collision in Apple NeuralHash model by arsome on Wednesday August 18, 2021 @08:57AM (#28220408)

To be fair, log_2(1 trillion) is only about 40 bits.

[ Parent | on HN ]