Results with Combined Noise lower than expected

[pierrefdz]

Hi and thanks for the code and architecture! I was wondering why the results with the combined noise are lower than the ones from the paper. Indeed, the original paper recovers 100% accuracy for the combined models when dealing with untransformed watermarked images, while we only obtain around 90% for networks trained with combined noise here. Do you happen to know why ?

[ando-khachatryan]

Hi, The noise layer implementation is a bit different (more general) from the paper. You can obtain the effect of the paper noise layers but you'd need to tweak with the noise layer parameters. See the description of noise layers and compare to the paper version for details.

[pierrefdz]

Hi and thanks for the quick answer. I've tried with dropout(0.3,0.3)+dropout(0.7,0.7)+cropout((0.84,0.84),(0.84,0.84))+cropout((0.55,0.55),(0.55,0.55))+crop((0.84,0.84),(0.84,0.84))+crop((0.55,0.55),(0.55,0.55))+blur(11,11)+blur(25,25)+jpeg() as noise, which should correspond to what is used in the paper, but I've obtained a BER of 0.1. However, I ran the experiment for more epochs and it seems that there is no convergence at 400 epochs, the issue might come from the fact that more iterations are needed than 400.