NeuralHash classifier

[kjsman]

I tried to build a classifier for NeuralHash: It gets NeuralHash as input and outputs class and probability.

I hashed all images of ILSVRC2012 dataset and trained the simple NN model.

Performance on the ImageNet validation dataset: (1,000 possible choices)

• Top-1 Accuracy: 5.25% (If random, 0.1% expected)
• Top-5 Accuracy: 14.09% (If random, 0.5% expected)

So... It seems that NeuralHash can't anonymize images well.

You can try this in Colab.

[willard-yuan]

@kjsman I think you should use the outs[0].flatten() as input, and NeuralHash is an embedding network which put out the outs[0].flatten() of 128 dimension. The seed.dot(outs[0].flatten()) is the Local Sensitive Hashing Process, and the seed is the projection matrix of LSH.

[kjsman]

@willard-yuan The goal of this project is to prove that NeuralHash cannot guarentee anonymity of image, so I used final form(including LSH) of NeuralHash because that is what Apple accesses.

[dxoigmn]

This is a nice datapoint. What was the simple NN model you used? I'm wondering if we could learn something about which bits of the hash are biased.