AppleNeuralHash2ONNX: Reverse-Engineered Apple NeuralHash, in ONNX and Python
As you may already know Apple is going to implement NeuralHash algorithm for on-device CSAM detection soon. Believe it or not, this algorithm already exists as early as iOS 14.3, hidden under obfuscated class names. After some digging and reverse engineering on the hidden APIs I managed to export its model (which is MobileNetV3) to ONNX and rebuild the whole NeuralHash algorithm in Python. You can now try NeuralHash even on Linux!
Source code: https://github.com/AsuharietYgvar/AppleNeuralHash2ONNX
No pre-exported model file will be provided here for obvious reasons. But it's very easy to export one yourself following the guide I included with the repo above. You don't even need any Apple devices to do it.
Early tests show that it can tolerate image resizing and compression, but not cropping or rotations.
Hope this will help us understand NeuralHash algorithm better and know its potential issues before it's enabled on all iOS devices.
Happy hacking!
Q: Can you explain more about how you know that the model extracted is the same NeuralHash that will be used for CSAM detection?
A: First of all, the model files have prefix NeuralHashv3b-, which is the same term as in Apple's document.
Secondly, in this document Apple described the algorithm details in Technology Overview -> NeuralHash section, which is exactly the same as what I discovered. For example, in Apple's document:
"Second, the descriptor is passed through a hashing scheme to convert the N floating-point numbers to M bits. Here, M is much smaller than the number of bits needed to represent the N floating-point numbers".
Оригинал:
https://www.reddit.com/r/MachineLearning/comments/p6hsoh/p_appleneuralhash2onnx_reverseengineered_apple/
Source code: https://github.com/AsuharietYgvar/AppleNeuralHash2ONNX
No pre-exported model file will be provided here for obvious reasons. But it's very easy to export one yourself following the guide I included with the repo above. You don't even need any Apple devices to do it.
Early tests show that it can tolerate image resizing and compression, but not cropping or rotations.
Hope this will help us understand NeuralHash algorithm better and know its potential issues before it's enabled on all iOS devices.
Happy hacking!
Q: Can you explain more about how you know that the model extracted is the same NeuralHash that will be used for CSAM detection?
A: First of all, the model files have prefix NeuralHashv3b-, which is the same term as in Apple's document.
Secondly, in this document Apple described the algorithm details in Technology Overview -> NeuralHash section, which is exactly the same as what I discovered. For example, in Apple's document:
"Second, the descriptor is passed through a hashing scheme to convert the N floating-point numbers to M bits. Here, M is much smaller than the number of bits needed to represent the N floating-point numbers".
Оригинал:
https://www.reddit.com/r/MachineLearning/comments/p6hsoh/p_appleneuralhash2onnx_reverseengineered_apple/