11–20 of 465 posts
I created a proof of concept showing how OpenAI's CLIP model can function as a "sanity check" similar to how Apple says their server-side model works. In order for a collision to get through to the human checkers, the same image would have to fool both networks independently: https://blog.roboflow.com/apples-csam-neuralhash-collision/
How long until thishashcollisionisnotporn.com is a thing?
For everyone upset about Apple's CSAM scanning, I think we all forgot about the EARN IT Act. It was nearly passed last year but Congress was finished before it could be voted on. It had Bipartisan support and would've virtually banned E2E of any kind. And it would have required scanning everywhere according to the recommendations of a 19-member board of NGOs and unelected experts. The reason for this mandatory backdo…
You can't ban encryption, it's practically impossible, it's like banning math.
> The system relies on a database of hashes—cryptographic representations of images—of known CSAM photos provided by National Center for Missing & Exploited Children (NCMEC) and other child protection organizations. “Cryptographic representations of images”. That’s not the case though right? These are “neuralhashes” afaik which are nowhere close to cryptographic hashes but rather locality sensitive hashes which is a…
I created a proof of concept showing how OpenAI's CLIP model can function as a "sanity check" similar to how Apple says their server-side model works. In order for a collision to get through to the human checkers, the same image would have to fool both networks independently: https://blog.roboflow.com/apples-csam-neuralhash-collision/
Cool project. Wouldn't padding all generated images with a few items that match CLIP get around this though?
Unclear how hard this would actually be in practice (if I were going to attempt it, the first thing I'd try is to evolve a colliding image with something like CLIP+VQGAN) but certainly harder than finding a collision alone.
> The system relies on a database of hashes—cryptographic representations of images—of known CSAM photos provided by National Center for Missing & Exploited Children (NCMEC) and other child protection organizations. “Cryptographic representations of images”. That’s not the case though right? These are “neuralhashes” afaik which are nowhere close to cryptographic hashes but rather locality sensitive hashes which is a…
Edit: "The main purpose of the hash is to ensure that identical and visually similar images result in the same hash, and images that are different from one another result in different hashes."[1] Apple isn't using a "similar image, similar hash" system. They're using a "similar image, same hash" system. [1]: https://www.apple.com/child-safety/pdf/CSAM_Detection_Techni...
Perceptual hashes are not cryptographic hashes. Perceptual hashing systems do compare hashes using a distance metric like the Hamming distance.
If two images have similar hashes, then they look kind of similar to one another. That's the point of perceptual hashing.
Hash collision in Apple NeuralHash model - https://news.ycombinator.com/item?id=28219068 - Aug 2021 (542 comments)
Convert Apple NeuralHash model for CSAM Detection to ONNX - https://news.ycombinator.com/item?id=28218391 - Aug 2021 (155 comments)
> The system relies on a database of hashes—cryptographic representations of images—of known CSAM photos provided by National Center for Missing & Exploited Children (NCMEC) and other child protection organizations. “Cryptographic representations of images”. That’s not the case though right? These are “neuralhashes” afaik which are nowhere close to cryptographic hashes but rather locality sensitive hashes which is a…
Edit: "The main purpose of the hash is to ensure that identical and visually similar images result in the same hash, and images that are different from one another result in different hashes."[1] Apple isn't using a "similar image, similar hash" system. They're using a "similar image, same hash" system. [1]: https://www.apple.com/child-safety/pdf/CSAM_Detection_Techni...
IGNORE THIS: I think that's the parent comment's point. These are definitely not cryptographic hashes, since they—by design and necessity—need to mirror hash similarity to the perceptual similarity of the input images.
For everyone upset about Apple's CSAM scanning, I think we all forgot about the EARN IT Act. It was nearly passed last year but Congress was finished before it could be voted on. It had Bipartisan support and would've virtually banned E2E of any kind. And it would have required scanning everywhere according to the recommendations of a 19-member board of NGOs and unelected experts. The reason for this mandatory backdo…
This would only make sense if Apple intends to expand their CSAM detection and reporting system to detect and report those other things, as well.
How long until thishashcollisionisnotporn.com is a thing?