What does HackerNews think of AppleNeuralHash2ONNX?

Convert Apple NeuralHash model for CSAM Detection to ONNX.

Language: Python

#16 in Machine learning
iCloud for Windows downloading other people's photos | Nov 2022
Expand Context ↕
Or just make your own fake/questionable hash collisions with a script Some Guy made on Github: https://github.com/AsuharietYgvar/AppleNeuralHash2ONNX
Apple Removes All References to Controversial CSAM Scanning Feature from Its… | Dec 2021
Expand Context ↕
Well, it’s mostly easy to tell since we still have researchers decompiling and scrubbing through the OS to see what’s in it and what it does[0].

https://github.com/AsuharietYgvar/AppleNeuralHash2ONNX
Show HN: Use your iPhone's camera to see any photo on your wall | Oct 2021
Expand Context ↕
I'm not big into "conspiracy" but you gotta wonder why this "NeuralHash" file on my computer exists if "NeuralHash" was supposedly delayed, according to Apple, on an undetermined timeline.

If I did want to cook up conspiracy theory, it would be easy: Apple wants to distract from the fact that NeuralHash was broken by researchers. This project apparently is able to create CSAM collisions:

https://github.com/AsuharietYgvar/AppleNeuralHash2ONNX

If NeuralHash is really rolled out, and if this python project can really create collisions, the CSAM system could be DDoS'd by people on their own computers, jamming up Apple's internal censorship review system with false positives. Hence, Apple would be incentivized to sweep this under the rug by "delaying" rollout indeterminately.
A catalog of naturally occurring images whose Apple NeuralHash is identical | Aug 2021
Why are exact collisions interesting? They are not intended to be compared exactly.

This algorithm doesn't even give exact matches for the same image on different hardware.

https://github.com/AsuharietYgvar/AppleNeuralHash2ONNX

Note: Neural hash generated here might be a few bits off from one generated on an iOS device. This is expected since different iOS devices generate slightly different hashes anyway. The reason is that neural networks are based on floating-point calculations. The accuracy is highly dependent on the hardware. For smaller networks it won't make any difference. But NeuralHash has 200+ layers, resulting in significant cumulative errors.

ImageNet contains naturally occurring Apple NeuralHash collisions | Aug 2021
Expand Context ↕
The instructions are here: https://github.com/AsuharietYgvar/AppleNeuralHash2ONNX

This was discussed on HN yesterday here (and a few other front-page stories): https://news.ycombinator.com/item?id=28218391