onnxjs VS AppleNeuralHash2ONNX

Tensorflow (and by extension Keras) offload most of the actual work to C++ or C, so having those compile to WebAssembly would (I imagine) be a herculean effort.

Instead, The TF team maintains TFJS, which can run on WebAssembly[0].

There are also tractjs[1], and onnyxjs[2], both of which allow you to run (most) ONNX models (which is an open standard for specifying ML models) using WebAssembly and WebGL(only onnyxjs supports WebGL). A bunch of frameworks (caffe, pytorch, TF) support exporting to/importing from ONNX.

[0] https://blog.tensorflow.org/2020/03/introducing-webassembly-...

[1] https://github.com/bminixhofer/tractjs

[2] https://github.com/microsoft/onnxjs

Perhaps you're referring to the hash collisions that were generated for the client-side NeuralHash which was reverse engineered about a year ago: https://github.com/AsuharietYgvar/AppleNeuralHash2ONNX/issues/1

Except Apple’s hashes have collisions where images that are similar in content but not identical will have the same hash. Meaning their process will have the same result: benign images being flagged as CSAM. See an example here https://github.com/AsuharietYgvar/AppleNeuralHash2ONNX/issues/1

perceptual hashes are completely broken due to false positives, like making a picture of a dog having the same hash as CSAM images. They are vulnerable to collision and have no preimage resistance, the most important feature. Because of that you can DOS conversations and people by sending dog images with the same hash as a CSAM: https://github.com/AsuharietYgvar/AppleNeuralHash2ONNX/issues/1

Specifically the fact that you can have one hash on two images, for example. There are other ways to abuse the CSAM system, which is confirmed by several researchers (one, two). And Apple saying it's not a concern... only to pull back on these plans.