Convert Apple NeuralHash model for CSAM Detection to ONNX.

Apple NeuralHash is a perceptual hashing method for images based on neural networks. It can tolerate image resize and compression. The steps of hashing is as the following:

1. Convert image to RGB.
2. Resize image to 360x360.
3. Normalize RGB values to [-1, 1] range.
4. Perform inference on the NeuralHash model.
5. Calculate dot product of a 96x128 matrix with the resulting vector of 128 floats.
6. Apply binary step to the resulting 96 float vector.
7. Convert the vector of 1.0 and 0.0 to bits, resulting in 96-bit binary data.

In this project, we convert Apple's NeuralHash model to ONNX format. A demo script for testing the model is also included.

Both macOS and Linux will work. In the following sections Debian is used for Linux example.

• macOS: Install by running brew install lzfse.
• Linux: Build and install from lzfse source.

Python 3.6 and above should work. Install the following dependencies:

pip install onnx coremltools

You will need 4 files from a recent macOS or iOS build:

• neuralhash_128x96_seed1.dat
• NeuralHashv3b-current.espresso.net
• NeuralHashv3b-current.espresso.shape
• NeuralHashv3b-current.espresso.weights

Option 1: From macOS or jailbroken iOS device (Recommended)

If you have a recent version of macOS (11.4+) or jailbroken iOS (14.7+) installed, simply grab these files from /System/Library/Frameworks/Vision.framework/Resources/ (on macOS) or /System/Library/Frameworks/Vision.framework/ (on iOS).

cd /path/to/ipsw/file
mkdir unpacked_ipsw
cd unpacked_ipsw
unzip ../*.ipsw

ls -lh

# Build and install apfs-fuse
sudo apt install fuse libfuse3-dev bzip2 libbz2-dev cmake g++ git libattr1-dev zlib1g-dev
git clone https://github.com/sgan81/apfs-fuse.git
cd apfs-fuse
git submodule init
git submodule update
mkdir build
cd build
cmake ..
make
sudo make install
sudo ln -s /bin/fusermount /bin/fusermount3
# Mount image
mkdir rootfs
apfs-fuse 018-63036-003.dmg rootfs

Put them under the same directory:

mkdir NeuralHash
cd NeuralHash
cp /System/Library/Frameworks/Vision.framework/Resources/NeuralHashv3b-current.espresso.* .
cp /System/Library/Frameworks/Vision.framework/Resources/neuralhash_128x96_seed1.dat .

Normally compiled Core ML models store structure in model.espresso.net and shapes in model.espresso.shape, both in JSON. It's the same for NeuralHash model but compressed with LZFSE.

dd if=NeuralHashv3b-current.espresso.net bs=4 skip=7 | lzfse -decode -o model.espresso.net
dd if=NeuralHashv3b-current.espresso.shape bs=4 skip=7 | lzfse -decode -o model.espresso.shape
cp NeuralHashv3b-current.espresso.weights model.espresso.weights

cd ..
git clone https://github.com/AsuharietYgvar/TNN.git
cd TNN
python3 tools/onnx2tnn/onnx-coreml/coreml2onnx.py ../NeuralHash

The resulting model is NeuralHash/model.onnx.

Netron is a perfect tool for this purpose.

1. Install required libraries:

pip install onnxruntime pillow

2. Run nnhash.py on an image:

python3 nnhash.py /path/to/model.onnx /path/to/neuralhash_128x96_seed1.dat image.jpg

Example output:

ab14febaa837b6c1484c35e6

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.

• nhcalc for uncovering NeuralHash private API.
• TNN for compiled Core ML to ONNX script.